Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * bufmgr.c
4 : : * buffer manager interface routines
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/storage/buffer/bufmgr.c
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : : /*
16 : : * Principal entry points:
17 : : *
18 : : * ReadBuffer() -- find or create a buffer holding the requested page,
19 : : * and pin it so that no one can destroy it while this process
20 : : * is using it.
21 : : *
22 : : * StartReadBuffer() -- as above, with separate wait step
23 : : * StartReadBuffers() -- multiple block version
24 : : * WaitReadBuffers() -- second step of above
25 : : *
26 : : * ReleaseBuffer() -- unpin a buffer
27 : : *
28 : : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
29 : : * The disk write is delayed until buffer replacement or checkpoint.
30 : : *
31 : : * See also these files:
32 : : * freelist.c -- chooses victim for buffer replacement
33 : : * buf_table.c -- manages the buffer lookup table
34 : : */
35 : : #include "postgres.h"
36 : :
37 : : #include <sys/file.h>
38 : : #include <unistd.h>
39 : :
40 : : #include "access/tableam.h"
41 : : #include "access/xloginsert.h"
42 : : #include "access/xlogutils.h"
43 : : #ifdef USE_ASSERT_CHECKING
44 : : #include "catalog/pg_tablespace_d.h"
45 : : #endif
46 : : #include "catalog/storage.h"
47 : : #include "catalog/storage_xlog.h"
48 : : #include "common/hashfn.h"
49 : : #include "executor/instrument.h"
50 : : #include "lib/binaryheap.h"
51 : : #include "miscadmin.h"
52 : : #include "pg_trace.h"
53 : : #include "pgstat.h"
54 : : #include "postmaster/bgwriter.h"
55 : : #include "storage/aio.h"
56 : : #include "storage/buf_internals.h"
57 : : #include "storage/bufmgr.h"
58 : : #include "storage/fd.h"
59 : : #include "storage/ipc.h"
60 : : #include "storage/lmgr.h"
61 : : #include "storage/proc.h"
62 : : #include "storage/proclist.h"
63 : : #include "storage/procsignal.h"
64 : : #include "storage/read_stream.h"
65 : : #include "storage/smgr.h"
66 : : #include "storage/standby.h"
67 : : #include "utils/memdebug.h"
68 : : #include "utils/ps_status.h"
69 : : #include "utils/rel.h"
70 : : #include "utils/resowner.h"
71 : : #include "utils/timestamp.h"
72 : : #include "utils/wait_event.h"
73 : :
74 : :
75 : : /* Note: these two macros only work on shared buffers, not local ones! */
76 : : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
77 : : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
78 : :
79 : : /* Note: this macro only works on local buffers, not shared ones! */
80 : : #define LocalBufHdrGetBlock(bufHdr) \
81 : : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
82 : :
83 : : /* Bits in SyncOneBuffer's return value */
84 : : #define BUF_WRITTEN 0x01
85 : : #define BUF_REUSABLE 0x02
86 : :
87 : : #define RELS_BSEARCH_THRESHOLD 20
88 : :
89 : : /*
90 : : * This is the size (in the number of blocks) above which we scan the
91 : : * entire buffer pool to remove the buffers for all the pages of relation
92 : : * being dropped. For the relations with size below this threshold, we find
93 : : * the buffers by doing lookups in BufMapping table.
94 : : */
95 : : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
96 : :
97 : : /*
98 : : * This is separated out from PrivateRefCountEntry to allow for copying all
99 : : * the data members via struct assignment.
100 : : */
101 : : typedef struct PrivateRefCountData
102 : : {
103 : : /*
104 : : * How many times has the buffer been pinned by this backend.
105 : : */
106 : : int32 refcount;
107 : :
108 : : /*
109 : : * Is the buffer locked by this backend? BUFFER_LOCK_UNLOCK indicates that
110 : : * the buffer is not locked.
111 : : */
112 : : BufferLockMode lockmode;
113 : : } PrivateRefCountData;
114 : :
115 : : typedef struct PrivateRefCountEntry
116 : : {
117 : : /*
118 : : * Note that this needs to be same as the entry's corresponding
119 : : * PrivateRefCountArrayKeys[i], if the entry is stored in the array. We
120 : : * store it in both places as this is used for the hashtable key and
121 : : * because it is more convenient (passing around a PrivateRefCountEntry
122 : : * suffices to identify the buffer) and faster (checking the keys array is
123 : : * faster when checking many entries, checking the entry is faster if just
124 : : * checking a single entry).
125 : : */
126 : : Buffer buffer;
127 : :
128 : : char status;
129 : :
130 : : PrivateRefCountData data;
131 : : } PrivateRefCountEntry;
132 : :
133 : : #define SH_PREFIX refcount
134 : : #define SH_ELEMENT_TYPE PrivateRefCountEntry
135 : : #define SH_KEY_TYPE Buffer
136 : : #define SH_KEY buffer
137 : : #define SH_HASH_KEY(tb, key) murmurhash32((uint32) (key))
138 : : #define SH_EQUAL(tb, a, b) ((a) == (b))
139 : : #define SH_SCOPE static inline
140 : : #define SH_DECLARE
141 : : #define SH_DEFINE
142 : : #include "lib/simplehash.h"
143 : :
144 : : /* 64 bytes, about the size of a cache line on common systems */
145 : : #define REFCOUNT_ARRAY_ENTRIES 8
146 : :
147 : : /*
148 : : * Status of buffers to checkpoint for a particular tablespace, used
149 : : * internally in BufferSync.
150 : : */
151 : : typedef struct CkptTsStatus
152 : : {
153 : : /* oid of the tablespace */
154 : : Oid tsId;
155 : :
156 : : /*
157 : : * Checkpoint progress for this tablespace. To make progress comparable
158 : : * between tablespaces the progress is, for each tablespace, measured as a
159 : : * number between 0 and the total number of to-be-checkpointed pages. Each
160 : : * page checkpointed in this tablespace increments this space's progress
161 : : * by progress_slice.
162 : : */
163 : : float8 progress;
164 : : float8 progress_slice;
165 : :
166 : : /* number of to-be checkpointed pages in this tablespace */
167 : : int num_to_scan;
168 : : /* already processed pages in this tablespace */
169 : : int num_scanned;
170 : :
171 : : /* current offset in CkptBufferIds for this tablespace */
172 : : int index;
173 : : } CkptTsStatus;
174 : :
175 : : /*
176 : : * Type for array used to sort SMgrRelations
177 : : *
178 : : * FlushRelationsAllBuffers shares the same comparator function with
179 : : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
180 : : * compatible.
181 : : */
182 : : typedef struct SMgrSortArray
183 : : {
184 : : RelFileLocator rlocator; /* This must be the first member */
185 : : SMgrRelation srel;
186 : : } SMgrSortArray;
187 : :
188 : : /* GUC variables */
189 : : bool zero_damaged_pages = false;
190 : : int bgwriter_lru_maxpages = 100;
191 : : double bgwriter_lru_multiplier = 2.0;
192 : : bool track_io_timing = false;
193 : :
194 : : /*
195 : : * How many buffers PrefetchBuffer callers should try to stay ahead of their
196 : : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
197 : : * for buffers not belonging to tablespaces that have their
198 : : * effective_io_concurrency parameter set.
199 : : */
200 : : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
201 : :
202 : : /*
203 : : * Like effective_io_concurrency, but used by maintenance code paths that might
204 : : * benefit from a higher setting because they work on behalf of many sessions.
205 : : * Overridden by the tablespace setting of the same name.
206 : : */
207 : : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
208 : :
209 : : /*
210 : : * Limit on how many blocks should be handled in single I/O operations.
211 : : * StartReadBuffers() callers should respect it, as should other operations
212 : : * that call smgr APIs directly. It is computed as the minimum of underlying
213 : : * GUCs io_combine_limit_guc and io_max_combine_limit.
214 : : */
215 : : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
216 : : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
217 : : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
218 : :
219 : : /*
220 : : * GUC variables about triggering kernel writeback for buffers written; OS
221 : : * dependent defaults are set via the GUC mechanism.
222 : : */
223 : : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
224 : : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
225 : : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
226 : :
227 : : /* local state for LockBufferForCleanup */
228 : : static BufferDesc *PinCountWaitBuf = NULL;
229 : :
230 : : /*
231 : : * Backend-Private refcount management:
232 : : *
233 : : * Each buffer also has a private refcount that keeps track of the number of
234 : : * times the buffer is pinned in the current process. This is so that the
235 : : * shared refcount needs to be modified only once if a buffer is pinned more
236 : : * than once by an individual backend. It's also used to check that no
237 : : * buffers are still pinned at the end of transactions and when exiting. We
238 : : * also use this mechanism to track whether this backend has a buffer locked,
239 : : * and, if so, in what mode.
240 : : *
241 : : *
242 : : * To avoid - as we used to - requiring an array with NBuffers entries to keep
243 : : * track of local buffers, we use a small sequentially searched array
244 : : * (PrivateRefCountArrayKeys, with the corresponding data stored in
245 : : * PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
246 : : * keep track of backend local pins.
247 : : *
248 : : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
249 : : * refcounts are kept track of in the array; after that, new array entries
250 : : * displace old ones into the hash table. That way a frequently used entry
251 : : * can't get "stuck" in the hashtable while infrequent ones clog the array.
252 : : *
253 : : * Note that in most scenarios the number of pinned buffers will not exceed
254 : : * REFCOUNT_ARRAY_ENTRIES.
255 : : *
256 : : *
257 : : * To enter a buffer into the refcount tracking mechanism first reserve a free
258 : : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
259 : : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
260 : : * memory allocations in NewPrivateRefCountEntry() which can be important
261 : : * because in some scenarios it's called with a spinlock held...
262 : : */
263 : : static Buffer PrivateRefCountArrayKeys[REFCOUNT_ARRAY_ENTRIES];
264 : : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
265 : : static refcount_hash *PrivateRefCountHash = NULL;
266 : : static int32 PrivateRefCountOverflowed = 0;
267 : : static uint32 PrivateRefCountClock = 0;
268 : : static int ReservedRefCountSlot = -1;
269 : : static int PrivateRefCountEntryLast = -1;
270 : :
271 : : static uint32 MaxProportionalPins;
272 : :
273 : : static void ReservePrivateRefCountEntry(void);
274 : : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
275 : : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
276 : : static inline int32 GetPrivateRefCount(Buffer buffer);
277 : : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
278 : :
279 : : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
280 : : static void ResOwnerReleaseBufferIO(Datum res);
281 : : static char *ResOwnerPrintBufferIO(Datum res);
282 : : static void ResOwnerReleaseBuffer(Datum res);
283 : : static char *ResOwnerPrintBuffer(Datum res);
284 : :
285 : : const ResourceOwnerDesc buffer_io_resowner_desc =
286 : : {
287 : : .name = "buffer io",
288 : : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
289 : : .release_priority = RELEASE_PRIO_BUFFER_IOS,
290 : : .ReleaseResource = ResOwnerReleaseBufferIO,
291 : : .DebugPrint = ResOwnerPrintBufferIO
292 : : };
293 : :
294 : : const ResourceOwnerDesc buffer_resowner_desc =
295 : : {
296 : : .name = "buffer",
297 : : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
298 : : .release_priority = RELEASE_PRIO_BUFFER_PINS,
299 : : .ReleaseResource = ResOwnerReleaseBuffer,
300 : : .DebugPrint = ResOwnerPrintBuffer
301 : : };
302 : :
303 : : /*
304 : : * Ensure that the PrivateRefCountArray has sufficient space to store one more
305 : : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
306 : : * a new entry - but it's perfectly fine to not use a reserved entry.
307 : : */
308 : : static void
309 : 89053843 : ReservePrivateRefCountEntry(void)
310 : : {
311 : : /* Already reserved (or freed), nothing to do */
312 [ + + ]: 89053843 : if (ReservedRefCountSlot != -1)
313 : 83806797 : return;
314 : :
315 : : /*
316 : : * First search for a free entry the array, that'll be sufficient in the
317 : : * majority of cases.
318 : : */
319 : : {
320 : : int i;
321 : :
322 [ + + ]: 47223414 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
323 : : {
324 [ + + ]: 41976368 : if (PrivateRefCountArrayKeys[i] == InvalidBuffer)
325 : : {
326 : 30709119 : ReservedRefCountSlot = i;
327 : :
328 : : /*
329 : : * We could return immediately, but iterating till the end of
330 : : * the array allows compiler-autovectorization.
331 : : */
332 : : }
333 : : }
334 : :
335 [ + + ]: 5247046 : if (ReservedRefCountSlot != -1)
336 : 5058235 : return;
337 : : }
338 : :
339 : : /*
340 : : * No luck. All array entries are full. Move one array entry into the hash
341 : : * table.
342 : : */
343 : : {
344 : : /*
345 : : * Move entry from the current clock position in the array into the
346 : : * hashtable. Use that slot.
347 : : */
348 : : int victim_slot;
349 : : PrivateRefCountEntry *victim_entry;
350 : : PrivateRefCountEntry *hashent;
351 : : bool found;
352 : :
353 : : /* select victim slot */
354 : 188811 : victim_slot = PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES;
355 : 188811 : victim_entry = &PrivateRefCountArray[victim_slot];
356 : 188811 : ReservedRefCountSlot = victim_slot;
357 : :
358 : : /* Better be used, otherwise we shouldn't get here. */
359 : : Assert(PrivateRefCountArrayKeys[victim_slot] != InvalidBuffer);
360 : : Assert(PrivateRefCountArray[victim_slot].buffer != InvalidBuffer);
361 : : Assert(PrivateRefCountArrayKeys[victim_slot] == PrivateRefCountArray[victim_slot].buffer);
362 : :
363 : : /* enter victim array entry into hashtable */
364 : 188811 : hashent = refcount_insert(PrivateRefCountHash,
365 : : PrivateRefCountArrayKeys[victim_slot],
366 : : &found);
367 : : Assert(!found);
368 : : /* move data from the entry in the array to the hash entry */
369 : 188811 : hashent->data = victim_entry->data;
370 : :
371 : : /* clear the now free array slot */
372 : 188811 : PrivateRefCountArrayKeys[victim_slot] = InvalidBuffer;
373 : 188811 : victim_entry->buffer = InvalidBuffer;
374 : :
375 : : /* clear the whole data member, just for future proofing */
376 : 188811 : memset(&victim_entry->data, 0, sizeof(victim_entry->data));
377 : 188811 : victim_entry->data.refcount = 0;
378 : 188811 : victim_entry->data.lockmode = BUFFER_LOCK_UNLOCK;
379 : :
380 : 188811 : PrivateRefCountOverflowed++;
381 : : }
382 : : }
383 : :
384 : : /*
385 : : * Fill a previously reserved refcount entry.
386 : : */
387 : : static PrivateRefCountEntry *
388 : 77357639 : NewPrivateRefCountEntry(Buffer buffer)
389 : : {
390 : : PrivateRefCountEntry *res;
391 : :
392 : : /* only allowed to be called when a reservation has been made */
393 : : Assert(ReservedRefCountSlot != -1);
394 : :
395 : : /* use up the reserved entry */
396 : 77357639 : res = &PrivateRefCountArray[ReservedRefCountSlot];
397 : :
398 : : /* and fill it */
399 : 77357639 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
400 : 77357639 : res->buffer = buffer;
401 : 77357639 : res->data.refcount = 0;
402 : 77357639 : res->data.lockmode = BUFFER_LOCK_UNLOCK;
403 : :
404 : : /* update cache for the next lookup */
405 : 77357639 : PrivateRefCountEntryLast = ReservedRefCountSlot;
406 : :
407 : 77357639 : ReservedRefCountSlot = -1;
408 : :
409 : 77357639 : return res;
410 : : }
411 : :
412 : : /*
413 : : * Slow-path for GetPrivateRefCountEntry(). This is big enough to not be worth
414 : : * inlining. This particularly seems to be true if the compiler is capable of
415 : : * auto-vectorizing the code, as that imposes additional stack-alignment
416 : : * requirements etc.
417 : : */
418 : : static pg_noinline PrivateRefCountEntry *
419 : 106884869 : GetPrivateRefCountEntrySlow(Buffer buffer, bool do_move)
420 : : {
421 : : PrivateRefCountEntry *res;
422 : 106884869 : int match = -1;
423 : : int i;
424 : :
425 : : /*
426 : : * First search for references in the array, that'll be sufficient in the
427 : : * majority of cases.
428 : : */
429 [ + + ]: 961963821 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
430 : : {
431 [ + + ]: 855078952 : if (PrivateRefCountArrayKeys[i] == buffer)
432 : : {
433 : 31827046 : match = i;
434 : : /* see ReservePrivateRefCountEntry() for why we don't return */
435 : : }
436 : : }
437 : :
438 [ + + ]: 106884869 : if (likely(match != -1))
439 : : {
440 : : /* update cache for the next lookup */
441 : 31827046 : PrivateRefCountEntryLast = match;
442 : :
443 : 31827046 : return &PrivateRefCountArray[match];
444 : : }
445 : :
446 : : /*
447 : : * By here we know that the buffer, if already pinned, isn't residing in
448 : : * the array.
449 : : *
450 : : * Only look up the buffer in the hashtable if we've previously overflowed
451 : : * into it.
452 : : */
453 [ + + ]: 75057823 : if (PrivateRefCountOverflowed == 0)
454 : 74401131 : return NULL;
455 : :
456 : 656692 : res = refcount_lookup(PrivateRefCountHash, buffer);
457 : :
458 [ + + ]: 656692 : if (res == NULL)
459 : 418054 : return NULL;
460 [ + + ]: 238638 : else if (!do_move)
461 : : {
462 : : /* caller doesn't want us to move the hash entry into the array */
463 : 141814 : return res;
464 : : }
465 : : else
466 : : {
467 : : /* move buffer from hashtable into the free array slot */
468 : : PrivateRefCountEntry *free;
469 : : PrivateRefCountData data;
470 : :
471 : : /* Save data and delete from hashtable while res is still valid */
472 : 96824 : data = res->data;
473 : 96824 : refcount_delete_item(PrivateRefCountHash, res);
474 : : Assert(PrivateRefCountOverflowed > 0);
475 : 96824 : PrivateRefCountOverflowed--;
476 : :
477 : : /* Ensure there's a free array slot */
478 : 96824 : ReservePrivateRefCountEntry();
479 : :
480 : : /* Use up the reserved slot */
481 : : Assert(ReservedRefCountSlot != -1);
482 : 96824 : free = &PrivateRefCountArray[ReservedRefCountSlot];
483 : : Assert(PrivateRefCountArrayKeys[ReservedRefCountSlot] == free->buffer);
484 : : Assert(free->buffer == InvalidBuffer);
485 : :
486 : : /* and fill it */
487 : 96824 : free->buffer = buffer;
488 : 96824 : free->data = data;
489 : 96824 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
490 : : /* update cache for the next lookup */
491 : 96824 : PrivateRefCountEntryLast = ReservedRefCountSlot;
492 : :
493 : 96824 : ReservedRefCountSlot = -1;
494 : :
495 : 96824 : return free;
496 : : }
497 : : }
498 : :
499 : : /*
500 : : * Return the PrivateRefCount entry for the passed buffer.
501 : : *
502 : : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
503 : : * do_move is true, and the entry resides in the hashtable the entry is
504 : : * optimized for frequent access by moving it to the array.
505 : : */
506 : : static inline PrivateRefCountEntry *
507 : 443248385 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
508 : : {
509 : : Assert(BufferIsValid(buffer));
510 : : Assert(!BufferIsLocal(buffer));
511 : :
512 : : /*
513 : : * It's very common to look up the same buffer repeatedly. To make that
514 : : * fast, we have a one-entry cache.
515 : : *
516 : : * In contrast to the loop in GetPrivateRefCountEntrySlow(), here it
517 : : * faster to check PrivateRefCountArray[].buffer, as in the case of a hit
518 : : * fewer addresses are computed and fewer cachelines are accessed. Whereas
519 : : * in GetPrivateRefCountEntrySlow()'s case, checking
520 : : * PrivateRefCountArrayKeys saves a lot of memory accesses.
521 : : */
522 [ + + ]: 443248385 : if (likely(PrivateRefCountEntryLast != -1) &&
523 [ + + ]: 443232087 : likely(PrivateRefCountArray[PrivateRefCountEntryLast].buffer == buffer))
524 : : {
525 : 336363516 : return &PrivateRefCountArray[PrivateRefCountEntryLast];
526 : : }
527 : :
528 : : /*
529 : : * The code for the cached lookup is small enough to be worth inlining
530 : : * into the caller. In the miss case however, that empirically doesn't
531 : : * seem worth it.
532 : : */
533 : 106884869 : return GetPrivateRefCountEntrySlow(buffer, do_move);
534 : : }
535 : :
536 : : /*
537 : : * Returns how many times the passed buffer is pinned by this backend.
538 : : *
539 : : * Only works for shared memory buffers!
540 : : */
541 : : static inline int32
542 : 3312443 : GetPrivateRefCount(Buffer buffer)
543 : : {
544 : : PrivateRefCountEntry *ref;
545 : :
546 : : Assert(BufferIsValid(buffer));
547 : : Assert(!BufferIsLocal(buffer));
548 : :
549 : : /*
550 : : * Not moving the entry - that's ok for the current users, but we might
551 : : * want to change this one day.
552 : : */
553 : 3312443 : ref = GetPrivateRefCountEntry(buffer, false);
554 : :
555 [ + + ]: 3312443 : if (ref == NULL)
556 : 30 : return 0;
557 : 3312413 : return ref->data.refcount;
558 : : }
559 : :
560 : : /*
561 : : * Release resources used to track the reference count of a buffer which we no
562 : : * longer have pinned and don't want to pin again immediately.
563 : : */
564 : : static void
565 : 77357639 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
566 : : {
567 : : Assert(ref->data.refcount == 0);
568 : : Assert(ref->data.lockmode == BUFFER_LOCK_UNLOCK);
569 : :
570 [ + - + + ]: 77357639 : if (ref >= &PrivateRefCountArray[0] &&
571 : : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
572 : : {
573 : 77265652 : ref->buffer = InvalidBuffer;
574 : 77265652 : PrivateRefCountArrayKeys[ref - PrivateRefCountArray] = InvalidBuffer;
575 : :
576 : :
577 : : /*
578 : : * Mark the just used entry as reserved - in many scenarios that
579 : : * allows us to avoid ever having to search the array/hash for free
580 : : * entries.
581 : : */
582 : 77265652 : ReservedRefCountSlot = ref - PrivateRefCountArray;
583 : : }
584 : : else
585 : : {
586 : 91987 : refcount_delete_item(PrivateRefCountHash, ref);
587 : : Assert(PrivateRefCountOverflowed > 0);
588 : 91987 : PrivateRefCountOverflowed--;
589 : : }
590 : 77357639 : }
591 : :
592 : : /*
593 : : * BufferIsPinned
594 : : * True iff the buffer is pinned (also checks for valid buffer number).
595 : : *
596 : : * NOTE: what we check here is that *this* backend holds a pin on
597 : : * the buffer. We do not care whether some other backend does.
598 : : */
599 : : #define BufferIsPinned(bufnum) \
600 : : ( \
601 : : !BufferIsValid(bufnum) ? \
602 : : false \
603 : : : \
604 : : BufferIsLocal(bufnum) ? \
605 : : (LocalRefCount[-(bufnum) - 1] > 0) \
606 : : : \
607 : : (GetPrivateRefCount(bufnum) > 0) \
608 : : )
609 : :
610 : :
611 : : static Buffer ReadBuffer_common(Relation rel,
612 : : SMgrRelation smgr, char smgr_persistence,
613 : : ForkNumber forkNum, BlockNumber blockNum,
614 : : ReadBufferMode mode, BufferAccessStrategy strategy);
615 : : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
616 : : ForkNumber fork,
617 : : BufferAccessStrategy strategy,
618 : : uint32 flags,
619 : : uint32 extend_by,
620 : : BlockNumber extend_upto,
621 : : Buffer *buffers,
622 : : uint32 *extended_by);
623 : : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
624 : : ForkNumber fork,
625 : : BufferAccessStrategy strategy,
626 : : uint32 flags,
627 : : uint32 extend_by,
628 : : BlockNumber extend_upto,
629 : : Buffer *buffers,
630 : : uint32 *extended_by);
631 : : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
632 : : bool skip_if_not_valid);
633 : : static void PinBuffer_Locked(BufferDesc *buf);
634 : : static void UnpinBuffer(BufferDesc *buf);
635 : : static void UnpinBufferNoOwner(BufferDesc *buf);
636 : : static void BufferSync(int flags);
637 : : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
638 : : WritebackContext *wb_context);
639 : : static void WaitIO(BufferDesc *buf);
640 : : static void AbortBufferIO(Buffer buffer);
641 : : static void shared_buffer_write_error_callback(void *arg);
642 : : static void local_buffer_write_error_callback(void *arg);
643 : : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
644 : : char relpersistence,
645 : : ForkNumber forkNum,
646 : : BlockNumber blockNum,
647 : : BufferAccessStrategy strategy,
648 : : bool *foundPtr, IOContext io_context);
649 : : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
650 : : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
651 : :
652 : : static pg_always_inline void TrackBufferHit(IOObject io_object,
653 : : IOContext io_context,
654 : : Relation rel, char persistence, SMgrRelation smgr,
655 : : ForkNumber forknum, BlockNumber blocknum);
656 : : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
657 : : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
658 : : IOObject io_object, IOContext io_context);
659 : : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
660 : : IOObject io_object, IOContext io_context);
661 : : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
662 : : ForkNumber forkNum,
663 : : BlockNumber nForkBlock,
664 : : BlockNumber firstDelBlock);
665 : : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
666 : : RelFileLocator dstlocator,
667 : : ForkNumber forkNum, bool permanent);
668 : : static void AtProcExit_Buffers(int code, Datum arg);
669 : : static void CheckForBufferLeaks(void);
670 : : #ifdef USE_ASSERT_CHECKING
671 : : static void AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode);
672 : : #endif
673 : : static int rlocator_comparator(const void *p1, const void *p2);
674 : : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
675 : : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
676 : : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
677 : :
678 : : static void BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
679 : : static void BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr);
680 : : static bool BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
681 : : static bool BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode);
682 : : static bool BufferLockHeldByMe(BufferDesc *buf_hdr);
683 : : static inline void BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr);
684 : : static inline int BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr);
685 : : static inline bool BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode);
686 : : static void BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode);
687 : : static void BufferLockDequeueSelf(BufferDesc *buf_hdr);
688 : : static void BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive);
689 : : static void BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate);
690 : : static inline uint64 BufferLockReleaseSub(BufferLockMode mode);
691 : :
692 : :
693 : : /*
694 : : * Implementation of PrefetchBuffer() for shared buffers.
695 : : */
696 : : PrefetchBufferResult
697 : 39980 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
698 : : ForkNumber forkNum,
699 : : BlockNumber blockNum)
700 : : {
701 : 39980 : PrefetchBufferResult result = {InvalidBuffer, false};
702 : : BufferTag newTag; /* identity of requested block */
703 : : uint32 newHash; /* hash value for newTag */
704 : : LWLock *newPartitionLock; /* buffer partition lock for it */
705 : : int buf_id;
706 : :
707 : : Assert(BlockNumberIsValid(blockNum));
708 : :
709 : : /* create a tag so we can lookup the buffer */
710 : 39980 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
711 : : forkNum, blockNum);
712 : :
713 : : /* determine its hash code and partition lock ID */
714 : 39980 : newHash = BufTableHashCode(&newTag);
715 : 39980 : newPartitionLock = BufMappingPartitionLock(newHash);
716 : :
717 : : /* see if the block is in the buffer pool already */
718 : 39980 : LWLockAcquire(newPartitionLock, LW_SHARED);
719 : 39980 : buf_id = BufTableLookup(&newTag, newHash);
720 : 39980 : LWLockRelease(newPartitionLock);
721 : :
722 : : /* If not in buffers, initiate prefetch */
723 [ + + ]: 39980 : if (buf_id < 0)
724 : : {
725 : : #ifdef USE_PREFETCH
726 : : /*
727 : : * Try to initiate an asynchronous read. This returns false in
728 : : * recovery if the relation file doesn't exist.
729 : : */
730 [ + + + - ]: 18885 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
731 : 9316 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
732 : : {
733 : 9316 : result.initiated_io = true;
734 : : }
735 : : #endif /* USE_PREFETCH */
736 : : }
737 : : else
738 : : {
739 : : /*
740 : : * Report the buffer it was in at that time. The caller may be able
741 : : * to avoid a buffer table lookup, but it's not pinned and it must be
742 : : * rechecked!
743 : : */
744 : 30411 : result.recent_buffer = buf_id + 1;
745 : : }
746 : :
747 : : /*
748 : : * If the block *is* in buffers, we do nothing. This is not really ideal:
749 : : * the block might be just about to be evicted, which would be stupid
750 : : * since we know we are going to need it soon. But the only easy answer
751 : : * is to bump the usage_count, which does not seem like a great solution:
752 : : * when the caller does ultimately touch the block, usage_count would get
753 : : * bumped again, resulting in too much favoritism for blocks that are
754 : : * involved in a prefetch sequence. A real fix would involve some
755 : : * additional per-buffer state, and it's not clear that there's enough of
756 : : * a problem to justify that.
757 : : */
758 : :
759 : 39980 : return result;
760 : : }
761 : :
762 : : /*
763 : : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
764 : : *
765 : : * This is named by analogy to ReadBuffer but doesn't actually allocate a
766 : : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
767 : : * block will not be delayed by the I/O. Prefetching is optional.
768 : : *
769 : : * There are three possible outcomes:
770 : : *
771 : : * 1. If the block is already cached, the result includes a valid buffer that
772 : : * could be used by the caller to avoid the need for a later buffer lookup, but
773 : : * it's not pinned, so the caller must recheck it.
774 : : *
775 : : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
776 : : * true. Currently there is no way to know if the data was already cached by
777 : : * the kernel and therefore didn't really initiate I/O, and no way to know when
778 : : * the I/O completes other than using synchronous ReadBuffer().
779 : : *
780 : : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
781 : : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
782 : : * lack of a kernel facility), direct I/O is enabled, or the underlying
783 : : * relation file wasn't found and we are in recovery. (If the relation file
784 : : * wasn't found and we are not in recovery, an error is raised).
785 : : */
786 : : PrefetchBufferResult
787 : 28995 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
788 : : {
789 : : Assert(RelationIsValid(reln));
790 : : Assert(BlockNumberIsValid(blockNum));
791 : :
792 [ + + ]: 28995 : if (RelationUsesLocalBuffers(reln))
793 : : {
794 : : /* see comments in ReadBuffer_common */
795 [ + - - + ]: 1357 : if (RELATION_IS_OTHER_TEMP(reln))
796 [ # # ]: 0 : ereport(ERROR,
797 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
798 : : errmsg("cannot access temporary tables of other sessions")));
799 : :
800 : : /* pass it off to localbuf.c */
801 : 1357 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
802 : : }
803 : : else
804 : : {
805 : : /* pass it to the shared buffer version */
806 : 27638 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
807 : : }
808 : : }
809 : :
810 : : /*
811 : : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
812 : : *
813 : : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
814 : : * successful. Return true if the buffer is valid and still has the expected
815 : : * tag. In that case, the buffer is pinned and the usage count is bumped.
816 : : */
817 : : bool
818 : 5114 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
819 : : Buffer recent_buffer)
820 : : {
821 : : BufferDesc *bufHdr;
822 : : BufferTag tag;
823 : : uint64 buf_state;
824 : :
825 : : Assert(BufferIsValid(recent_buffer));
826 : :
827 : 5114 : ResourceOwnerEnlarge(CurrentResourceOwner);
828 : 5114 : ReservePrivateRefCountEntry();
829 : 5114 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
830 : :
831 [ + + ]: 5114 : if (BufferIsLocal(recent_buffer))
832 : : {
833 : 140 : int b = -recent_buffer - 1;
834 : :
835 : 140 : bufHdr = GetLocalBufferDescriptor(b);
836 : 140 : buf_state = pg_atomic_read_u64(&bufHdr->state);
837 : :
838 : : /* Is it still valid and holding the right tag? */
839 [ + - + - ]: 140 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
840 : : {
841 : 140 : PinLocalBuffer(bufHdr, true);
842 : :
843 : 140 : pgBufferUsage.local_blks_hit++;
844 : :
845 : 140 : return true;
846 : : }
847 : : }
848 : : else
849 : : {
850 : 4974 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
851 : :
852 : : /*
853 : : * Is it still valid and holding the right tag? We do an unlocked tag
854 : : * comparison first, to make it unlikely that we'll increment the
855 : : * usage counter of the wrong buffer, if someone calls us with a very
856 : : * out of date recent_buffer. Then we'll check it again if we get the
857 : : * pin.
858 : : */
859 [ + + + + ]: 9915 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
860 : 4941 : PinBuffer(bufHdr, NULL, true))
861 : : {
862 [ + - ]: 4935 : if (BufferTagsEqual(&tag, &bufHdr->tag))
863 : : {
864 : 4935 : pgBufferUsage.shared_blks_hit++;
865 : 4935 : return true;
866 : : }
867 : 0 : UnpinBuffer(bufHdr);
868 : : }
869 : : }
870 : :
871 : 39 : return false;
872 : : }
873 : :
874 : : /*
875 : : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
876 : : * fork with RBM_NORMAL mode and default strategy.
877 : : */
878 : : Buffer
879 : 63512762 : ReadBuffer(Relation reln, BlockNumber blockNum)
880 : : {
881 : 63512762 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
882 : : }
883 : :
884 : : /*
885 : : * ReadBufferExtended -- returns a buffer containing the requested
886 : : * block of the requested relation. If the blknum
887 : : * requested is P_NEW, extend the relation file and
888 : : * allocate a new block. (Caller is responsible for
889 : : * ensuring that only one backend tries to extend a
890 : : * relation at the same time!)
891 : : *
892 : : * Returns: the buffer number for the buffer containing
893 : : * the block read. The returned buffer has been pinned.
894 : : * Does not return on error --- elog's instead.
895 : : *
896 : : * Assume when this function is called, that reln has been opened already.
897 : : *
898 : : * In RBM_NORMAL mode, the page is read from disk, and the page header is
899 : : * validated. An error is thrown if the page header is not valid. (But
900 : : * note that an all-zero page is considered "valid"; see
901 : : * PageIsVerified().)
902 : : *
903 : : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
904 : : * valid, the page is zeroed instead of throwing an error. This is intended
905 : : * for non-critical data, where the caller is prepared to repair errors.
906 : : *
907 : : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
908 : : * filled with zeros instead of reading it from disk. Useful when the caller
909 : : * is going to fill the page from scratch, since this saves I/O and avoids
910 : : * unnecessary failure if the page-on-disk has corrupt page headers.
911 : : * The page is returned locked to ensure that the caller has a chance to
912 : : * initialize the page before it's made visible to others.
913 : : * Caution: do not use this mode to read a page that is beyond the relation's
914 : : * current physical EOF; that is likely to cause problems in md.c when
915 : : * the page is modified and written out. P_NEW is OK, though.
916 : : *
917 : : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
918 : : * a cleanup-strength lock on the page.
919 : : *
920 : : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
921 : : *
922 : : * If strategy is not NULL, a nondefault buffer access strategy is used.
923 : : * See buffer/README for details.
924 : : */
925 : : inline Buffer
926 : 75033713 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
927 : : ReadBufferMode mode, BufferAccessStrategy strategy)
928 : : {
929 : : Buffer buf;
930 : :
931 : : /*
932 : : * Read the buffer, and update pgstat counters to reflect a cache hit or
933 : : * miss. The other-session temp-relation check is enforced by
934 : : * ReadBuffer_common().
935 : : */
936 : 75033713 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
937 : : forkNum, blockNum, mode, strategy);
938 : :
939 : 75033688 : return buf;
940 : : }
941 : :
942 : :
943 : : /*
944 : : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
945 : : * a relcache entry for the relation.
946 : : *
947 : : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
948 : : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
949 : : * cannot be used for temporary relations (and making that work might be
950 : : * difficult, unless we only want to read temporary relations for our own
951 : : * ProcNumber).
952 : : */
953 : : Buffer
954 : 5995719 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
955 : : BlockNumber blockNum, ReadBufferMode mode,
956 : : BufferAccessStrategy strategy, bool permanent)
957 : : {
958 : 5995719 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
959 : :
960 [ + - ]: 5995719 : return ReadBuffer_common(NULL, smgr,
961 : : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
962 : : forkNum, blockNum,
963 : : mode, strategy);
964 : : }
965 : :
966 : : /*
967 : : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
968 : : */
969 : : Buffer
970 : 57449 : ExtendBufferedRel(BufferManagerRelation bmr,
971 : : ForkNumber forkNum,
972 : : BufferAccessStrategy strategy,
973 : : uint32 flags)
974 : : {
975 : : Buffer buf;
976 : 57449 : uint32 extend_by = 1;
977 : :
978 : 57449 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
979 : : &buf, &extend_by);
980 : :
981 : 57449 : return buf;
982 : : }
983 : :
984 : : /*
985 : : * Extend relation by multiple blocks.
986 : : *
987 : : * Tries to extend the relation by extend_by blocks. Depending on the
988 : : * availability of resources the relation may end up being extended by a
989 : : * smaller number of pages (unless an error is thrown, always by at least one
990 : : * page). *extended_by is updated to the number of pages the relation has been
991 : : * extended to.
992 : : *
993 : : * buffers needs to be an array that is at least extend_by long. Upon
994 : : * completion, the first extend_by array elements will point to a pinned
995 : : * buffer.
996 : : *
997 : : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
998 : : * locked. This is useful for callers that want a buffer that is guaranteed to
999 : : * be empty.
1000 : : */
1001 : : BlockNumber
1002 : 211697 : ExtendBufferedRelBy(BufferManagerRelation bmr,
1003 : : ForkNumber fork,
1004 : : BufferAccessStrategy strategy,
1005 : : uint32 flags,
1006 : : uint32 extend_by,
1007 : : Buffer *buffers,
1008 : : uint32 *extended_by)
1009 : : {
1010 : : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1011 : : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1012 : : Assert(extend_by > 0);
1013 : :
1014 [ + - ]: 211697 : if (bmr.relpersistence == '\0')
1015 : 211697 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1016 : :
1017 : 211697 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1018 : : extend_by, InvalidBlockNumber,
1019 : : buffers, extended_by);
1020 : : }
1021 : :
1022 : : /*
1023 : : * Extend the relation so it is at least extend_to blocks large, return buffer
1024 : : * (extend_to - 1).
1025 : : *
1026 : : * This is useful for callers that want to write a specific page, regardless
1027 : : * of the current size of the relation (e.g. useful for visibilitymap and for
1028 : : * crash recovery).
1029 : : */
1030 : : Buffer
1031 : 56030 : ExtendBufferedRelTo(BufferManagerRelation bmr,
1032 : : ForkNumber fork,
1033 : : BufferAccessStrategy strategy,
1034 : : uint32 flags,
1035 : : BlockNumber extend_to,
1036 : : ReadBufferMode mode)
1037 : : {
1038 : : BlockNumber current_size;
1039 : 56030 : uint32 extended_by = 0;
1040 : 56030 : Buffer buffer = InvalidBuffer;
1041 : : Buffer buffers[64];
1042 : :
1043 : : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1044 : : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1045 : : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
1046 : :
1047 [ + + ]: 56030 : if (bmr.relpersistence == '\0')
1048 : 9492 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1049 : :
1050 : : /*
1051 : : * If desired, create the file if it doesn't exist. If
1052 : : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
1053 : : * an smgrexists call.
1054 : : */
1055 [ + + ]: 56030 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
1056 [ + - + + ]: 9492 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
1057 [ + - - + ]: 31 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
1058 [ + - + + ]: 9461 : !smgrexists(BMR_GET_SMGR(bmr), fork))
1059 : : {
1060 : 9429 : LockRelationForExtension(bmr.rel, ExclusiveLock);
1061 : :
1062 : : /* recheck, fork might have been created concurrently */
1063 [ + - + + ]: 9429 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
1064 [ + - ]: 9428 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
1065 : :
1066 : 9429 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
1067 : : }
1068 : :
1069 : : /*
1070 : : * If requested, invalidate size cache, so that smgrnblocks asks the
1071 : : * kernel.
1072 : : */
1073 [ + + ]: 56030 : if (flags & EB_CLEAR_SIZE_CACHE)
1074 [ + - ]: 9492 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1075 : :
1076 : : /*
1077 : : * Estimate how many pages we'll need to extend by. This avoids acquiring
1078 : : * unnecessarily many victim buffers.
1079 : : */
1080 [ + + ]: 56030 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
1081 : :
1082 : : /*
1083 : : * Since no-one else can be looking at the page contents yet, there is no
1084 : : * difference between an exclusive lock and a cleanup-strength lock. Note
1085 : : * that we pass the original mode to ReadBuffer_common() below, when
1086 : : * falling back to reading the buffer to a concurrent relation extension.
1087 : : */
1088 [ + + - + ]: 56030 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1089 : 46140 : flags |= EB_LOCK_TARGET;
1090 : :
1091 [ + + ]: 114301 : while (current_size < extend_to)
1092 : : {
1093 : 58271 : uint32 num_pages = lengthof(buffers);
1094 : : BlockNumber first_block;
1095 : :
1096 [ + + ]: 58271 : if ((uint64) current_size + num_pages > extend_to)
1097 : 58205 : num_pages = extend_to - current_size;
1098 : :
1099 : 58271 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1100 : : num_pages, extend_to,
1101 : : buffers, &extended_by);
1102 : :
1103 : 58271 : current_size = first_block + extended_by;
1104 : : Assert(num_pages != 0 || current_size >= extend_to);
1105 : :
1106 [ + + ]: 125372 : for (uint32 i = 0; i < extended_by; i++)
1107 : : {
1108 [ + + ]: 67101 : if (first_block + i != extend_to - 1)
1109 : 11077 : ReleaseBuffer(buffers[i]);
1110 : : else
1111 : 56024 : buffer = buffers[i];
1112 : : }
1113 : : }
1114 : :
1115 : : /*
1116 : : * It's possible that another backend concurrently extended the relation.
1117 : : * In that case read the buffer.
1118 : : *
1119 : : * XXX: Should we control this via a flag?
1120 : : */
1121 [ + + ]: 56030 : if (buffer == InvalidBuffer)
1122 : : {
1123 : : Assert(extended_by == 0);
1124 [ + - ]: 6 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1125 : : fork, extend_to - 1, mode, strategy);
1126 : : }
1127 : :
1128 : 56030 : return buffer;
1129 : : }
1130 : :
1131 : : /*
1132 : : * Lock and optionally zero a buffer, as part of the implementation of
1133 : : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1134 : : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1135 : : */
1136 : : static void
1137 : 355806 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1138 : : {
1139 : : BufferDesc *bufHdr;
1140 : : bool need_to_zero;
1141 : 355806 : bool isLocalBuf = BufferIsLocal(buffer);
1142 : : StartBufferIOResult sbres;
1143 : :
1144 : : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1145 : :
1146 [ + + ]: 355806 : if (already_valid)
1147 : : {
1148 : : /*
1149 : : * If the caller already knew the buffer was valid, we can skip some
1150 : : * header interaction. The caller just wants to lock the buffer.
1151 : : */
1152 : 38578 : need_to_zero = false;
1153 : : }
1154 : : else
1155 : : {
1156 [ + + ]: 317228 : if (isLocalBuf)
1157 : : {
1158 : : /* Simple case for non-shared buffers. */
1159 : 30 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1160 : 30 : sbres = StartLocalBufferIO(bufHdr, true, true, NULL);
1161 : : }
1162 : : else
1163 : : {
1164 : : /*
1165 : : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1166 : : * concurrently. Even though we aren't doing I/O, that ensures
1167 : : * that we don't zero a page that someone else has pinned. An
1168 : : * exclusive content lock wouldn't be enough, because readers are
1169 : : * allowed to drop the content lock after determining that a tuple
1170 : : * is visible (see buffer access rules in README).
1171 : : */
1172 : 317198 : bufHdr = GetBufferDescriptor(buffer - 1);
1173 : 317198 : sbres = StartSharedBufferIO(bufHdr, true, true, NULL);
1174 : : }
1175 : :
1176 : : Assert(sbres != BUFFER_IO_IN_PROGRESS);
1177 : 317228 : need_to_zero = sbres == BUFFER_IO_READY_FOR_IO;
1178 : : }
1179 : :
1180 [ + + ]: 355806 : if (need_to_zero)
1181 : : {
1182 : 317228 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1183 : :
1184 : : /*
1185 : : * Grab the buffer content lock before marking the page as valid, to
1186 : : * make sure that no other backend sees the zeroed page before the
1187 : : * caller has had a chance to initialize it.
1188 : : *
1189 : : * Since no-one else can be looking at the page contents yet, there is
1190 : : * no difference between an exclusive lock and a cleanup-strength
1191 : : * lock. (Note that we cannot use LockBuffer() or
1192 : : * LockBufferForCleanup() here, because they assert that the buffer is
1193 : : * already valid.)
1194 : : */
1195 [ + + ]: 317228 : if (!isLocalBuf)
1196 : 317198 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1197 : :
1198 : : /* Set BM_VALID, terminate IO, and wake up any waiters */
1199 [ + + ]: 317228 : if (isLocalBuf)
1200 : 30 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1201 : : else
1202 : 317198 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1203 : : }
1204 [ + + ]: 38578 : else if (!isLocalBuf)
1205 : : {
1206 : : /*
1207 : : * The buffer is valid, so we can't zero it. The caller still expects
1208 : : * the page to be locked on return.
1209 : : */
1210 [ + + ]: 38558 : if (mode == RBM_ZERO_AND_LOCK)
1211 : 38488 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1212 : : else
1213 : 70 : LockBufferForCleanup(buffer);
1214 : : }
1215 : 355806 : }
1216 : :
1217 : : /*
1218 : : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1219 : : * already present, or false if more work is required to either read it in or
1220 : : * zero it.
1221 : : */
1222 : : static pg_always_inline Buffer
1223 : 85787139 : PinBufferForBlock(Relation rel,
1224 : : SMgrRelation smgr,
1225 : : char persistence,
1226 : : ForkNumber forkNum,
1227 : : BlockNumber blockNum,
1228 : : BufferAccessStrategy strategy,
1229 : : IOObject io_object,
1230 : : IOContext io_context,
1231 : : bool *foundPtr)
1232 : : {
1233 : : BufferDesc *bufHdr;
1234 : :
1235 : : Assert(blockNum != P_NEW);
1236 : :
1237 : : /* Persistence should be set before */
1238 : : Assert((persistence == RELPERSISTENCE_TEMP ||
1239 : : persistence == RELPERSISTENCE_PERMANENT ||
1240 : : persistence == RELPERSISTENCE_UNLOGGED));
1241 : :
1242 : : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1243 : : smgr->smgr_rlocator.locator.spcOid,
1244 : : smgr->smgr_rlocator.locator.dbOid,
1245 : : smgr->smgr_rlocator.locator.relNumber,
1246 : : smgr->smgr_rlocator.backend);
1247 : :
1248 [ + + ]: 85787139 : if (persistence == RELPERSISTENCE_TEMP)
1249 : 1648797 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1250 : : else
1251 : 84138342 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1252 : : strategy, foundPtr, io_context);
1253 : :
1254 [ + + ]: 85787131 : if (*foundPtr)
1255 : 83886008 : TrackBufferHit(io_object, io_context, rel, persistence, smgr, forkNum, blockNum);
1256 : :
1257 [ + + ]: 85787131 : if (rel)
1258 : : {
1259 : : /*
1260 : : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1261 : : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1262 : : * zeroed instead), the per-relation stats always count them.
1263 : : */
1264 [ + + + + : 79534062 : pgstat_count_buffer_read(rel);
+ + ]
1265 : : }
1266 : :
1267 : 85787131 : return BufferDescriptorGetBuffer(bufHdr);
1268 : : }
1269 : :
1270 : : /*
1271 : : * ReadBuffer_common -- common logic for all ReadBuffer variants
1272 : : *
1273 : : * smgr is required, rel is optional unless using P_NEW.
1274 : : */
1275 : : static pg_always_inline Buffer
1276 : 81029438 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1277 : : ForkNumber forkNum,
1278 : : BlockNumber blockNum, ReadBufferMode mode,
1279 : : BufferAccessStrategy strategy)
1280 : : {
1281 : : ReadBuffersOperation operation;
1282 : : Buffer buffer;
1283 : : int flags;
1284 : : char persistence;
1285 : :
1286 : : /*
1287 : : * Reject attempts to read non-local temporary relations; we would be
1288 : : * likely to get wrong data since we have no visibility into the owning
1289 : : * session's local buffers. This is the canonical place for the check,
1290 : : * covering the ReadBufferExtended() entry point and any other caller that
1291 : : * supplies a Relation.
1292 : : */
1293 [ + + + + : 81029438 : if (rel && RELATION_IS_OTHER_TEMP(rel))
+ + ]
1294 [ + - ]: 2 : ereport(ERROR,
1295 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1296 : : errmsg("cannot access temporary tables of other sessions")));
1297 : :
1298 : : /*
1299 : : * Backward compatibility path, most code should use ExtendBufferedRel()
1300 : : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1301 : : * scales a lot better.
1302 : : */
1303 [ + + ]: 81029436 : if (unlikely(blockNum == P_NEW))
1304 : : {
1305 : 322 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1306 : :
1307 : : /*
1308 : : * Since no-one else can be looking at the page contents yet, there is
1309 : : * no difference between an exclusive lock and a cleanup-strength
1310 : : * lock.
1311 : : */
1312 [ + - - + ]: 322 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1313 : 0 : flags |= EB_LOCK_FIRST;
1314 : :
1315 : 322 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1316 : : }
1317 : :
1318 [ + + ]: 81029114 : if (rel)
1319 : 75033395 : persistence = rel->rd_rel->relpersistence;
1320 : : else
1321 : 5995719 : persistence = smgr_persistence;
1322 : :
1323 [ + + + + : 81029114 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
+ + ]
1324 : : mode == RBM_ZERO_AND_LOCK))
1325 : : {
1326 : : bool found;
1327 : : IOContext io_context;
1328 : : IOObject io_object;
1329 : :
1330 [ + + ]: 355806 : if (persistence == RELPERSISTENCE_TEMP)
1331 : : {
1332 : 50 : io_context = IOCONTEXT_NORMAL;
1333 : 50 : io_object = IOOBJECT_TEMP_RELATION;
1334 : : }
1335 : : else
1336 : : {
1337 : 355756 : io_context = IOContextForStrategy(strategy);
1338 : 355756 : io_object = IOOBJECT_RELATION;
1339 : : }
1340 : :
1341 : 355806 : buffer = PinBufferForBlock(rel, smgr, persistence,
1342 : : forkNum, blockNum, strategy,
1343 : : io_object, io_context, &found);
1344 : 355806 : ZeroAndLockBuffer(buffer, mode, found);
1345 : 355806 : return buffer;
1346 : : }
1347 : :
1348 : : /*
1349 : : * Signal that we are going to immediately wait. If we're immediately
1350 : : * waiting, there is no benefit in actually executing the IO
1351 : : * asynchronously, it would just add dispatch overhead.
1352 : : */
1353 : 80673308 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1354 [ + + ]: 80673308 : if (mode == RBM_ZERO_ON_ERROR)
1355 : 2027239 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1356 : 80673308 : operation.smgr = smgr;
1357 : 80673308 : operation.rel = rel;
1358 : 80673308 : operation.persistence = persistence;
1359 : 80673308 : operation.forknum = forkNum;
1360 : 80673308 : operation.strategy = strategy;
1361 [ + + ]: 80673308 : if (StartReadBuffer(&operation,
1362 : : &buffer,
1363 : : blockNum,
1364 : : flags))
1365 : 772551 : WaitReadBuffers(&operation);
1366 : :
1367 : 80673285 : return buffer;
1368 : : }
1369 : :
1370 : : static pg_always_inline bool
1371 : 85247212 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1372 : : Buffer *buffers,
1373 : : BlockNumber blockNum,
1374 : : int *nblocks,
1375 : : int flags,
1376 : : bool allow_forwarding)
1377 : : {
1378 : 85247212 : int actual_nblocks = *nblocks;
1379 : 85247212 : int maxcombine = 0;
1380 : : bool did_start_io;
1381 : : IOContext io_context;
1382 : : IOObject io_object;
1383 : :
1384 : : Assert(*nblocks == 1 || allow_forwarding);
1385 : : Assert(*nblocks > 0);
1386 : : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1387 : :
1388 : : /* see comments in ReadBuffer_common */
1389 [ + + + + : 85247212 : if (operation->rel && RELATION_IS_OTHER_TEMP(operation->rel))
- + ]
1390 [ # # ]: 0 : ereport(ERROR,
1391 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1392 : : errmsg("cannot access temporary tables of other sessions")));
1393 : :
1394 [ + + ]: 85247212 : if (operation->persistence == RELPERSISTENCE_TEMP)
1395 : : {
1396 : 1640551 : io_context = IOCONTEXT_NORMAL;
1397 : 1640551 : io_object = IOOBJECT_TEMP_RELATION;
1398 : : }
1399 : : else
1400 : : {
1401 : 83606661 : io_context = IOContextForStrategy(operation->strategy);
1402 : 83606661 : io_object = IOOBJECT_RELATION;
1403 : : }
1404 : :
1405 [ + + ]: 86831121 : for (int i = 0; i < actual_nblocks; ++i)
1406 : : {
1407 : : bool found;
1408 : :
1409 [ + + + + ]: 85433541 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1410 : 2208 : {
1411 : : BufferDesc *bufHdr;
1412 : :
1413 : : /*
1414 : : * This is a buffer that was pinned by an earlier call to
1415 : : * StartReadBuffers(), but couldn't be handled in one operation at
1416 : : * that time. The operation was split, and the caller has passed
1417 : : * an already pinned buffer back to us to handle the rest of the
1418 : : * operation. It must continue at the expected block number.
1419 : : */
1420 : : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1421 : :
1422 : : /*
1423 : : * It might be an already valid buffer (a hit) that followed the
1424 : : * final contiguous block of an earlier I/O (a miss) marking the
1425 : : * end of it, or a buffer that some other backend has since made
1426 : : * valid by performing the I/O for us, in which case we can handle
1427 : : * it as a hit now. It is safe to check for a BM_VALID flag with
1428 : : * a relaxed load, because we got a fresh view of it while pinning
1429 : : * it in the previous call.
1430 : : *
1431 : : * On the other hand if we don't see BM_VALID yet, it must be an
1432 : : * I/O that was split by the previous call and we need to try to
1433 : : * start a new I/O from this block. We're also racing against any
1434 : : * other backend that might start the I/O or even manage to mark
1435 : : * it BM_VALID after this check, but StartBufferIO() will handle
1436 : : * those cases.
1437 : : */
1438 [ + + ]: 2208 : if (BufferIsLocal(buffers[i]))
1439 : 16 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1440 : : else
1441 : 2192 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1442 : : Assert(pg_atomic_read_u64(&bufHdr->state) & BM_TAG_VALID);
1443 : 2208 : found = pg_atomic_read_u64(&bufHdr->state) & BM_VALID;
1444 : : }
1445 : : else
1446 : : {
1447 : 85431325 : buffers[i] = PinBufferForBlock(operation->rel,
1448 : : operation->smgr,
1449 : 85431333 : operation->persistence,
1450 : : operation->forknum,
1451 : : blockNum + i,
1452 : : operation->strategy,
1453 : : io_object, io_context,
1454 : : &found);
1455 : : }
1456 : :
1457 [ + + ]: 85433533 : if (found)
1458 : : {
1459 : : /*
1460 : : * We have a hit. If it's the first block in the requested range,
1461 : : * we can return it immediately and report that WaitReadBuffers()
1462 : : * does not need to be called. If the initial value of *nblocks
1463 : : * was larger, the caller will have to call again for the rest.
1464 : : */
1465 [ + + ]: 83849624 : if (i == 0)
1466 : : {
1467 : 83847428 : *nblocks = 1;
1468 : :
1469 : : #ifdef USE_ASSERT_CHECKING
1470 : :
1471 : : /*
1472 : : * Initialize enough of ReadBuffersOperation to make
1473 : : * CheckReadBuffersOperation() work. Outside of assertions
1474 : : * that's not necessary when no IO is issued.
1475 : : */
1476 : : operation->buffers = buffers;
1477 : : operation->blocknum = blockNum;
1478 : : operation->nblocks = 1;
1479 : : operation->nblocks_done = 1;
1480 : : CheckReadBuffersOperation(operation, true);
1481 : : #endif
1482 : 83847428 : return false;
1483 : : }
1484 : :
1485 : : /*
1486 : : * Otherwise we already have an I/O to perform, but this block
1487 : : * can't be included as it is already valid. Split the I/O here.
1488 : : * There may or may not be more blocks requiring I/O after this
1489 : : * one, we haven't checked, but they can't be contiguous with this
1490 : : * one in the way. We'll leave this buffer pinned, forwarding it
1491 : : * to the next call, avoiding the need to unpin it here and re-pin
1492 : : * it in the next call.
1493 : : */
1494 : 2196 : actual_nblocks = i;
1495 : 2196 : break;
1496 : : }
1497 : : else
1498 : : {
1499 : : /*
1500 : : * Check how many blocks we can cover with the same IO. The smgr
1501 : : * implementation might e.g. be limited due to a segment boundary.
1502 : : */
1503 [ + + + + ]: 1583909 : if (i == 0 && actual_nblocks > 1)
1504 : : {
1505 : 37858 : maxcombine = smgrmaxcombine(operation->smgr,
1506 : : operation->forknum,
1507 : : blockNum);
1508 [ - + ]: 37858 : if (unlikely(maxcombine < actual_nblocks))
1509 : : {
1510 [ # # ]: 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1511 : : blockNum, actual_nblocks, maxcombine);
1512 : 0 : actual_nblocks = maxcombine;
1513 : : }
1514 : : }
1515 : : }
1516 : : }
1517 : 1399776 : *nblocks = actual_nblocks;
1518 : :
1519 : : /* Populate information needed for I/O. */
1520 : 1399776 : operation->buffers = buffers;
1521 : 1399776 : operation->blocknum = blockNum;
1522 : 1399776 : operation->flags = flags;
1523 : 1399776 : operation->nblocks = actual_nblocks;
1524 : 1399776 : operation->nblocks_done = 0;
1525 : 1399776 : pgaio_wref_clear(&operation->io_wref);
1526 : :
1527 : : /*
1528 : : * When using AIO, start the IO in the background. If not, issue prefetch
1529 : : * requests if desired by the caller.
1530 : : *
1531 : : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1532 : : * de-risk the introduction of AIO somewhat. It's a large architectural
1533 : : * change, with lots of chances for unanticipated performance effects.
1534 : : *
1535 : : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1536 : : * asynchronously, but without the check here we'd execute IO earlier than
1537 : : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1538 : : */
1539 [ + + ]: 1399776 : if (io_method != IOMETHOD_SYNC)
1540 : : {
1541 : : /*
1542 : : * Try to start IO asynchronously. It's possible that no IO needs to
1543 : : * be started, if another backend already performed the IO.
1544 : : *
1545 : : * Note that if an IO is started, it might not cover the entire
1546 : : * requested range, e.g. because an intermediary block has been read
1547 : : * in by another backend. In that case any "trailing" buffers we
1548 : : * already pinned above will be "forwarded" by read_stream.c to the
1549 : : * next call to StartReadBuffers().
1550 : : *
1551 : : * This is signalled to the caller by decrementing *nblocks *and*
1552 : : * reducing operation->nblocks. The latter is done here, but not below
1553 : : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1554 : : * overall read size anymore, we need to retry until done in its
1555 : : * entirety or until failed.
1556 : : */
1557 : 1398393 : did_start_io = AsyncReadBuffers(operation, nblocks);
1558 : :
1559 : 1398378 : operation->nblocks = *nblocks;
1560 : : }
1561 : : else
1562 : : {
1563 : 1383 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1564 : :
1565 [ + + ]: 1383 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1566 : : {
1567 : : /*
1568 : : * In theory we should only do this if PinBufferForBlock() had to
1569 : : * allocate new buffers above. That way, if two calls to
1570 : : * StartReadBuffers() were made for the same blocks before
1571 : : * WaitReadBuffers(), only the first would issue the advice.
1572 : : * That'd be a better simulation of true asynchronous I/O, which
1573 : : * would only start the I/O once, but isn't done here for
1574 : : * simplicity.
1575 : : */
1576 : 19 : smgrprefetch(operation->smgr,
1577 : : operation->forknum,
1578 : : blockNum,
1579 : : actual_nblocks);
1580 : : }
1581 : :
1582 : : /*
1583 : : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1584 : : * will initiate the necessary IO.
1585 : : */
1586 : 1383 : did_start_io = true;
1587 : : }
1588 : :
1589 : 1399761 : CheckReadBuffersOperation(operation, !did_start_io);
1590 : :
1591 : 1399761 : return did_start_io;
1592 : : }
1593 : :
1594 : : /*
1595 : : * Begin reading a range of blocks beginning at blockNum and extending for
1596 : : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1597 : : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1598 : : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1599 : : * and is now being continued. On return, *nblocks holds the number of blocks
1600 : : * accepted by this operation. If it is less than the original number then
1601 : : * this operation has been split, but buffer elements up to the original
1602 : : * requested size may hold forwarded buffers to be used for a continuing
1603 : : * operation. The caller must either start a new I/O beginning at the block
1604 : : * immediately following the blocks accepted by this call and pass those
1605 : : * buffers back in, or release them if it chooses not to. It shouldn't make
1606 : : * any other use of or assumptions about forwarded buffers.
1607 : : *
1608 : : * If false is returned, no I/O is necessary and the buffers covered by
1609 : : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1610 : : * an I/O has been started, and WaitReadBuffers() must be called with the same
1611 : : * operation object before the buffers covered by *nblocks on exit can be
1612 : : * accessed. Along with the operation object, the caller-supplied array of
1613 : : * buffers must remain valid until WaitReadBuffers() is called, and any
1614 : : * forwarded buffers must also be preserved for a continuing call unless
1615 : : * they are explicitly released.
1616 : : */
1617 : : bool
1618 : 2108478 : StartReadBuffers(ReadBuffersOperation *operation,
1619 : : Buffer *buffers,
1620 : : BlockNumber blockNum,
1621 : : int *nblocks,
1622 : : int flags)
1623 : : {
1624 : 2108478 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1625 : : true /* expect forwarded buffers */ );
1626 : : }
1627 : :
1628 : : /*
1629 : : * Single block version of the StartReadBuffers(). This might save a few
1630 : : * instructions when called from another translation unit, because it is
1631 : : * specialized for nblocks == 1.
1632 : : *
1633 : : * This version does not support "forwarded" buffers: they cannot be created
1634 : : * by reading only one block and *buffer is ignored on entry.
1635 : : */
1636 : : bool
1637 : 83138734 : StartReadBuffer(ReadBuffersOperation *operation,
1638 : : Buffer *buffer,
1639 : : BlockNumber blocknum,
1640 : : int flags)
1641 : : {
1642 : 83138734 : int nblocks = 1;
1643 : : bool result;
1644 : :
1645 : 83138734 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1646 : : false /* single block, no forwarding */ );
1647 : : Assert(nblocks == 1); /* single block can't be short */
1648 : :
1649 : 83138719 : return result;
1650 : : }
1651 : :
1652 : : /*
1653 : : * Perform sanity checks on the ReadBuffersOperation.
1654 : : */
1655 : : static void
1656 : 4199321 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1657 : : {
1658 : : #ifdef USE_ASSERT_CHECKING
1659 : : Assert(operation->nblocks_done <= operation->nblocks);
1660 : : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1661 : :
1662 : : for (int i = 0; i < operation->nblocks; i++)
1663 : : {
1664 : : Buffer buffer = operation->buffers[i];
1665 : : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1666 : : GetLocalBufferDescriptor(-buffer - 1) :
1667 : : GetBufferDescriptor(buffer - 1);
1668 : :
1669 : : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1670 : : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_TAG_VALID);
1671 : :
1672 : : if (i < operation->nblocks_done)
1673 : : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_VALID);
1674 : : }
1675 : : #endif
1676 : 4199321 : }
1677 : :
1678 : : /*
1679 : : * We track various stats related to buffer hits. Because this is done in a
1680 : : * few separate places, this helper exists for convenience.
1681 : : */
1682 : : static pg_always_inline void
1683 : 83888344 : TrackBufferHit(IOObject io_object, IOContext io_context,
1684 : : Relation rel, char persistence, SMgrRelation smgr,
1685 : : ForkNumber forknum, BlockNumber blocknum)
1686 : : {
1687 : : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum,
1688 : : blocknum,
1689 : : smgr->smgr_rlocator.locator.spcOid,
1690 : : smgr->smgr_rlocator.locator.dbOid,
1691 : : smgr->smgr_rlocator.locator.relNumber,
1692 : : smgr->smgr_rlocator.backend,
1693 : : true);
1694 : :
1695 [ + + ]: 83888344 : if (persistence == RELPERSISTENCE_TEMP)
1696 : 1637744 : pgBufferUsage.local_blks_hit += 1;
1697 : : else
1698 : 82250600 : pgBufferUsage.shared_blks_hit += 1;
1699 : :
1700 : 83888344 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1701 : :
1702 [ + + ]: 83888344 : if (VacuumCostActive)
1703 : 2978392 : VacuumCostBalance += VacuumCostPageHit;
1704 : :
1705 [ + + ]: 83888344 : if (rel)
1706 [ + + + + : 78152375 : pgstat_count_buffer_hit(rel);
+ + ]
1707 : 83888344 : }
1708 : :
1709 : : /*
1710 : : * Helper for WaitReadBuffers() that processes the results of a readv
1711 : : * operation, raising an error if necessary.
1712 : : */
1713 : : static void
1714 : 1397141 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1715 : : {
1716 : 1397141 : PgAioReturn *aio_ret = &operation->io_return;
1717 : 1397141 : PgAioResultStatus rs = aio_ret->result.status;
1718 : 1397141 : int newly_read_blocks = 0;
1719 : :
1720 : : Assert(pgaio_wref_valid(&operation->io_wref));
1721 : : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1722 : :
1723 : : /*
1724 : : * SMGR reports the number of blocks successfully read as the result of
1725 : : * the IO operation. Thus we can simply add that to ->nblocks_done.
1726 : : */
1727 : :
1728 [ + + ]: 1397141 : if (likely(rs != PGAIO_RS_ERROR))
1729 : 1397112 : newly_read_blocks = aio_ret->result.result;
1730 : :
1731 [ + + + + ]: 1397141 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1732 [ + + ]: 46 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1733 : : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1734 [ + + ]: 1397095 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1735 : : {
1736 : : /*
1737 : : * We'll retry, so we just emit a debug message to the server log (or
1738 : : * not even that in prod scenarios).
1739 : : */
1740 : 109 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1741 [ + - ]: 109 : elog(DEBUG3, "partial read, will retry");
1742 : : }
1743 : :
1744 : : Assert(newly_read_blocks > 0);
1745 : : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1746 : :
1747 : 1397112 : operation->nblocks_done += newly_read_blocks;
1748 : :
1749 : : Assert(operation->nblocks_done <= operation->nblocks);
1750 : 1397112 : }
1751 : :
1752 : : /*
1753 : : * Wait for the IO operation initiated by StartReadBuffers() et al to
1754 : : * complete.
1755 : : *
1756 : : * Returns true if we needed to wait for the IO operation, false otherwise.
1757 : : */
1758 : : bool
1759 : 1399046 : WaitReadBuffers(ReadBuffersOperation *operation)
1760 : : {
1761 : 1399046 : PgAioReturn *aio_ret = &operation->io_return;
1762 : : IOContext io_context;
1763 : : IOObject io_object;
1764 : 1399046 : bool needed_wait = false;
1765 : :
1766 [ + + ]: 1399046 : if (operation->persistence == RELPERSISTENCE_TEMP)
1767 : : {
1768 : 2429 : io_context = IOCONTEXT_NORMAL;
1769 : 2429 : io_object = IOOBJECT_TEMP_RELATION;
1770 : : }
1771 : : else
1772 : : {
1773 : 1396617 : io_context = IOContextForStrategy(operation->strategy);
1774 : 1396617 : io_object = IOOBJECT_RELATION;
1775 : : }
1776 : :
1777 : : /*
1778 : : * If we get here without an IO operation having been issued, the
1779 : : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1780 : : * caller should not have called WaitReadBuffers().
1781 : : *
1782 : : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1783 : : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1784 : : * of the retry logic below, no extra code is required.
1785 : : *
1786 : : * This path is expected to eventually go away.
1787 : : */
1788 [ + + + - ]: 1399046 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1789 [ # # ]: 0 : elog(ERROR, "waiting for read operation that didn't read");
1790 : :
1791 : : /*
1792 : : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1793 : : * done. We may need multiple retries, not just because we could get
1794 : : * multiple partial reads, but also because some of the remaining
1795 : : * to-be-read buffers may have been read in by other backends, limiting
1796 : : * the IO size.
1797 : : */
1798 : : while (true)
1799 : 1497 : {
1800 : : int ignored_nblocks_progress;
1801 : :
1802 : 1400543 : CheckReadBuffersOperation(operation, false);
1803 : :
1804 : : /*
1805 : : * If there is an IO associated with the operation, we may need to
1806 : : * wait for it.
1807 : : */
1808 [ + + ]: 1400543 : if (pgaio_wref_valid(&operation->io_wref))
1809 : : {
1810 : : /*
1811 : : * Track the time spent waiting for the IO to complete. As
1812 : : * tracking a wait even if we don't actually need to wait
1813 : : *
1814 : : * a) is not cheap, due to the timestamping overhead
1815 : : *
1816 : : * b) reports some time as waiting, even if we never waited
1817 : : *
1818 : : * we first check if we already know the IO is complete.
1819 : : *
1820 : : * Note that operation->io_return is uninitialized for foreign IO,
1821 : : * so we cannot use the cheaper PGAIO_RS_UNKNOWN pre-check.
1822 : : */
1823 [ + + + + ]: 1399151 : if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
1824 [ + + ]: 604428 : !pgaio_wref_check_done(&operation->io_wref))
1825 : : {
1826 : 298012 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1827 : :
1828 : 298012 : pgaio_wref_wait(&operation->io_wref);
1829 : 298012 : needed_wait = true;
1830 : :
1831 : : /*
1832 : : * The IO operation itself was already counted earlier, in
1833 : : * AsyncReadBuffers(), this just accounts for the wait time.
1834 : : */
1835 : 298012 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1836 : : io_start, 0, 0);
1837 : : }
1838 : : else
1839 : : {
1840 : : Assert(pgaio_wref_check_done(&operation->io_wref));
1841 : : }
1842 : :
1843 [ + + ]: 1399151 : if (unlikely(operation->foreign_io))
1844 : : {
1845 : 2010 : Buffer buffer = operation->buffers[operation->nblocks_done];
1846 : 2010 : BufferDesc *desc = BufferIsLocal(buffer) ?
1847 [ - + ]: 2010 : GetLocalBufferDescriptor(-buffer - 1) :
1848 : 2010 : GetBufferDescriptor(buffer - 1);
1849 : 2010 : uint64 buf_state = pg_atomic_read_u64(&desc->state);
1850 : :
1851 [ + + ]: 2010 : if (buf_state & BM_VALID)
1852 : : {
1853 : 2008 : BlockNumber blocknum = operation->blocknum + operation->nblocks_done;
1854 : :
1855 : 2008 : operation->nblocks_done += 1;
1856 : : Assert(operation->nblocks_done <= operation->nblocks);
1857 : :
1858 : : /*
1859 : : * Track this as a 'hit' for this backend. The backend
1860 : : * performing the IO will track it as a 'read'.
1861 : : */
1862 : 2008 : TrackBufferHit(io_object, io_context,
1863 : 2008 : operation->rel, operation->persistence,
1864 : : operation->smgr, operation->forknum,
1865 : : blocknum);
1866 : : }
1867 : :
1868 : : /*
1869 : : * If the foreign IO failed and left the buffer invalid,
1870 : : * nblocks_done is not incremented. The retry loop below will
1871 : : * call AsyncReadBuffers() which will attempt the IO itself.
1872 : : */
1873 : : }
1874 : : else
1875 : : {
1876 : : /*
1877 : : * We now are sure the IO completed. Check the results. This
1878 : : * includes reporting on errors if there were any.
1879 : : */
1880 : 1397141 : ProcessReadBuffersResult(operation);
1881 : : }
1882 : : }
1883 : :
1884 : : /*
1885 : : * Most of the time, the one IO we already started, will read in
1886 : : * everything. But we need to deal with partial reads and buffers not
1887 : : * needing IO anymore.
1888 : : */
1889 [ + + ]: 1400514 : if (operation->nblocks_done == operation->nblocks)
1890 : 1399017 : break;
1891 : :
1892 [ - + ]: 1497 : CHECK_FOR_INTERRUPTS();
1893 : :
1894 : : /*
1895 : : * If the IO completed only partially, we need to perform additional
1896 : : * work, consider that a form of having had to wait.
1897 : : */
1898 : 1497 : needed_wait = true;
1899 : :
1900 : : /*
1901 : : * This may only complete the IO partially, either because some
1902 : : * buffers were already valid, or because of a partial read.
1903 : : *
1904 : : * NB: In contrast to after the AsyncReadBuffers() call in
1905 : : * StartReadBuffers(), we do *not* reduce
1906 : : * ReadBuffersOperation->nblocks here, callers expect the full
1907 : : * operation to be completed at this point (as more operations may
1908 : : * have been queued).
1909 : : */
1910 : 1497 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1911 : : }
1912 : :
1913 : 1399017 : CheckReadBuffersOperation(operation, true);
1914 : :
1915 : : /* NB: READ_DONE tracepoint was already executed in completion callback */
1916 : 1399017 : return needed_wait;
1917 : : }
1918 : :
1919 : : /*
1920 : : * Initiate IO for the ReadBuffersOperation
1921 : : *
1922 : : * This function only starts a single IO at a time. The size of the IO may be
1923 : : * limited to below the to-be-read blocks, if one of the buffers has
1924 : : * concurrently been read in. If the first to-be-read buffer is already valid,
1925 : : * no IO will be issued.
1926 : : *
1927 : : * To support retries after partial reads, the first operation->nblocks_done
1928 : : * buffers are skipped.
1929 : : *
1930 : : * On return *nblocks_progress is updated to reflect the number of buffers
1931 : : * affected by the call. If the first buffer is valid, *nblocks_progress is
1932 : : * set to 1 and operation->nblocks_done is incremented.
1933 : : *
1934 : : * Returns true if IO was initiated or is already in progress (foreign IO),
1935 : : * false if the buffer was already valid.
1936 : : */
1937 : : static bool
1938 : 1399890 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1939 : : {
1940 : 1399890 : Buffer *buffers = &operation->buffers[0];
1941 : 1399890 : int flags = operation->flags;
1942 : 1399890 : ForkNumber forknum = operation->forknum;
1943 : 1399890 : char persistence = operation->persistence;
1944 : 1399890 : int16 nblocks_done = operation->nblocks_done;
1945 : 1399890 : BlockNumber blocknum = operation->blocknum + nblocks_done;
1946 : 1399890 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1947 : 1399890 : int io_buffers_len = 0;
1948 : : PgAioHandle *ioh;
1949 : 1399890 : uint32 ioh_flags = 0;
1950 : : void *io_pages[MAX_IO_COMBINE_LIMIT];
1951 : : IOContext io_context;
1952 : : IOObject io_object;
1953 : : instr_time io_start;
1954 : : StartBufferIOResult status;
1955 : :
1956 [ + + ]: 1399890 : if (persistence == RELPERSISTENCE_TEMP)
1957 : : {
1958 : 2821 : io_context = IOCONTEXT_NORMAL;
1959 : 2821 : io_object = IOOBJECT_TEMP_RELATION;
1960 : : }
1961 : : else
1962 : : {
1963 : 1397069 : io_context = IOContextForStrategy(operation->strategy);
1964 : 1397069 : io_object = IOOBJECT_RELATION;
1965 : : }
1966 : :
1967 : : /*
1968 : : * When this IO is executed synchronously, either because the caller will
1969 : : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1970 : : * the AIO subsystem needs to know.
1971 : : */
1972 [ + + ]: 1399890 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1973 : 786810 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1974 : :
1975 [ + + ]: 1399890 : if (persistence == RELPERSISTENCE_TEMP)
1976 : 2821 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1977 : :
1978 : : /*
1979 : : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1980 : : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1981 : : * set globally, but on a per-session basis. The completion callback,
1982 : : * which may be run in other processes, e.g. in IO workers, may have a
1983 : : * different value of the zero_damaged_pages GUC.
1984 : : *
1985 : : * XXX: We probably should eventually use a different flag for
1986 : : * zero_damaged_pages, so we can report different log levels / error codes
1987 : : * for zero_damaged_pages and ZERO_ON_ERROR.
1988 : : */
1989 [ + + ]: 1399890 : if (zero_damaged_pages)
1990 : 16 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1991 : :
1992 : : /*
1993 : : * For the same reason as with zero_damaged_pages we need to use this
1994 : : * backend's ignore_checksum_failure value.
1995 : : */
1996 [ + + ]: 1399890 : if (ignore_checksum_failure)
1997 : 8 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1998 : :
1999 : :
2000 : : /*
2001 : : * To be allowed to report stats in the local completion callback we need
2002 : : * to prepare to report stats now. This ensures we can safely report the
2003 : : * checksum failure even in a critical section.
2004 : : */
2005 : 1399890 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
2006 : :
2007 : : /*
2008 : : * We must get an IO handle before StartBufferIO(), as pgaio_io_acquire()
2009 : : * might block, which we don't want after setting IO_IN_PROGRESS. If we
2010 : : * don't need to do the IO, we'll release the handle.
2011 : : *
2012 : : * If we need to wait for IO before we can get a handle, submit
2013 : : * already-staged IO first, so that other backends don't need to wait.
2014 : : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
2015 : : * wait for already submitted IO, which doesn't require additional locks,
2016 : : * but it could still cause undesirable waits.
2017 : : *
2018 : : * A secondary benefit is that this would allow us to measure the time in
2019 : : * pgaio_io_acquire() without causing undue timer overhead in the common,
2020 : : * non-blocking, case. However, currently the pgstats infrastructure
2021 : : * doesn't really allow that, as it a) asserts that an operation can't
2022 : : * have time without operations b) doesn't have an API to report
2023 : : * "accumulated" time.
2024 : : */
2025 : 1399890 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
2026 [ + + ]: 1399890 : if (unlikely(!ioh))
2027 : : {
2028 : 3120 : pgaio_submit_staged();
2029 : 3120 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
2030 : : }
2031 : :
2032 : 1399890 : operation->foreign_io = false;
2033 : 1399890 : pgaio_wref_clear(&operation->io_wref);
2034 : :
2035 : : /*
2036 : : * Try to start IO on the first buffer in a new run of blocks. If AIO is
2037 : : * in progress, be it in this backend or another backend, we just
2038 : : * associate the wait reference with the operation and wait in
2039 : : * WaitReadBuffers(). This turns out to be important for performance in
2040 : : * two workloads:
2041 : : *
2042 : : * 1) A read stream that has to read the same block multiple times within
2043 : : * the readahead distance. This can happen e.g. for the table accesses of
2044 : : * an index scan.
2045 : : *
2046 : : * 2) Concurrent scans by multiple backends on the same relation.
2047 : : *
2048 : : * If we were to synchronously wait for the in-progress IO, we'd not be
2049 : : * able to keep enough I/O in flight.
2050 : : *
2051 : : * If we do find there is ongoing I/O for the buffer, we set up a 1-block
2052 : : * ReadBuffersOperation that WaitReadBuffers then can wait on.
2053 : : *
2054 : : * It's possible that another backend has started IO on the buffer but not
2055 : : * yet set its wait reference. In this case, we have no choice but to wait
2056 : : * for either the wait reference to be valid or the IO to be done.
2057 : : */
2058 : 1399890 : status = StartBufferIO(buffers[nblocks_done], true, true,
2059 : : &operation->io_wref);
2060 [ + + ]: 1399890 : if (status != BUFFER_IO_READY_FOR_IO)
2061 : : {
2062 : 2338 : pgaio_io_release(ioh);
2063 : 2338 : *nblocks_progress = 1;
2064 [ + + ]: 2338 : if (status == BUFFER_IO_ALREADY_DONE)
2065 : : {
2066 : : /*
2067 : : * Someone has already completed this block, we're done.
2068 : : *
2069 : : * When IO is necessary, ->nblocks_done is updated in
2070 : : * ProcessReadBuffersResult(), but that is not called if no IO is
2071 : : * necessary. Thus update here.
2072 : : */
2073 : 328 : operation->nblocks_done += 1;
2074 : : Assert(operation->nblocks_done <= operation->nblocks);
2075 : :
2076 : : Assert(!pgaio_wref_valid(&operation->io_wref));
2077 : :
2078 : : /*
2079 : : * Report and track this as a 'hit' for this backend, even though
2080 : : * it must have started out as a miss in PinBufferForBlock(). The
2081 : : * other backend will track this as a 'read'.
2082 : : */
2083 : 328 : TrackBufferHit(io_object, io_context,
2084 : 328 : operation->rel, operation->persistence,
2085 : : operation->smgr, operation->forknum,
2086 : : blocknum);
2087 : 328 : return false;
2088 : : }
2089 : :
2090 : : /* The IO is already in-progress */
2091 : : Assert(status == BUFFER_IO_IN_PROGRESS);
2092 : : Assert(pgaio_wref_valid(&operation->io_wref));
2093 : 2010 : operation->foreign_io = true;
2094 : :
2095 : 2010 : return true;
2096 : : }
2097 : :
2098 : : Assert(io_buffers[0] == buffers[nblocks_done]);
2099 : 1397552 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
2100 : 1397552 : io_buffers_len = 1;
2101 : :
2102 : : /*
2103 : : * NB: As little code as possible should be added between the
2104 : : * StartBufferIO() above, the further StartBufferIO()s below and the
2105 : : * smgrstartreadv(), as some of the buffers are now marked as
2106 : : * IO_IN_PROGRESS and will thus cause other backends to wait.
2107 : : */
2108 : :
2109 : : /*
2110 : : * How many neighboring-on-disk blocks can we scatter-read into other
2111 : : * buffers at the same time? In this case we don't wait if we see an I/O
2112 : : * already in progress (see comment above).
2113 : : */
2114 [ + + ]: 1582318 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
2115 : : {
2116 : : /* Must be consecutive block numbers. */
2117 : : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
2118 : : BufferGetBlockNumber(buffers[i]) - 1);
2119 : :
2120 : 184770 : status = StartBufferIO(buffers[i], true, false, NULL);
2121 [ + + ]: 184770 : if (status != BUFFER_IO_READY_FOR_IO)
2122 : 4 : break;
2123 : :
2124 : : Assert(io_buffers[io_buffers_len] == buffers[i]);
2125 : :
2126 : 184766 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
2127 : : }
2128 : :
2129 : : /* get a reference to wait for in WaitReadBuffers() */
2130 : 1397552 : pgaio_io_get_wref(ioh, &operation->io_wref);
2131 : :
2132 : : /* provide the list of buffers to the completion callbacks */
2133 : 1397552 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
2134 : :
2135 [ + + ]: 1397552 : pgaio_io_register_callbacks(ioh,
2136 : : persistence == RELPERSISTENCE_TEMP ?
2137 : : PGAIO_HCB_LOCAL_BUFFER_READV :
2138 : : PGAIO_HCB_SHARED_BUFFER_READV,
2139 : : flags);
2140 : :
2141 : 1397552 : pgaio_io_set_flag(ioh, ioh_flags);
2142 : :
2143 : : /* ---
2144 : : * Even though we're trying to issue IO asynchronously, track the time
2145 : : * in smgrstartreadv():
2146 : : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
2147 : : * immediately
2148 : : * - the io method might not support the IO (e.g. worker IO for a temp
2149 : : * table)
2150 : : * ---
2151 : : */
2152 : 1397552 : io_start = pgstat_prepare_io_time(track_io_timing);
2153 : 1397552 : smgrstartreadv(ioh, operation->smgr, forknum,
2154 : : blocknum,
2155 : : io_pages, io_buffers_len);
2156 : 1397537 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
2157 : 1397537 : io_start, 1, io_buffers_len * BLCKSZ);
2158 : :
2159 [ + + ]: 1397537 : if (persistence == RELPERSISTENCE_TEMP)
2160 : 2819 : pgBufferUsage.local_blks_read += io_buffers_len;
2161 : : else
2162 : 1394718 : pgBufferUsage.shared_blks_read += io_buffers_len;
2163 : :
2164 : : /*
2165 : : * Track vacuum cost when issuing IO, not after waiting for it. Otherwise
2166 : : * we could end up issuing a lot of IO in a short timespan, despite a low
2167 : : * cost limit.
2168 : : */
2169 [ + + ]: 1397537 : if (VacuumCostActive)
2170 : 21345 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
2171 : :
2172 : 1397537 : *nblocks_progress = io_buffers_len;
2173 : :
2174 : 1397537 : return true;
2175 : : }
2176 : :
2177 : : /*
2178 : : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
2179 : : * buffer. If no buffer exists already, selects a replacement victim and
2180 : : * evicts the old page, but does NOT read in new page.
2181 : : *
2182 : : * "strategy" can be a buffer replacement strategy object, or NULL for
2183 : : * the default strategy. The selected buffer's usage_count is advanced when
2184 : : * using the default strategy, but otherwise possibly not (see PinBuffer).
2185 : : *
2186 : : * The returned buffer is pinned and is already marked as holding the
2187 : : * desired page. If it already did have the desired page, *foundPtr is
2188 : : * set true. Otherwise, *foundPtr is set false.
2189 : : *
2190 : : * io_context is passed as an output parameter to avoid calling
2191 : : * IOContextForStrategy() when there is a shared buffers hit and no IO
2192 : : * statistics need be captured.
2193 : : *
2194 : : * No locks are held either at entry or exit.
2195 : : */
2196 : : static pg_always_inline BufferDesc *
2197 : 84138342 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2198 : : BlockNumber blockNum,
2199 : : BufferAccessStrategy strategy,
2200 : : bool *foundPtr, IOContext io_context)
2201 : : {
2202 : : BufferTag newTag; /* identity of requested block */
2203 : : uint32 newHash; /* hash value for newTag */
2204 : : LWLock *newPartitionLock; /* buffer partition lock for it */
2205 : : int existing_buf_id;
2206 : : Buffer victim_buffer;
2207 : : BufferDesc *victim_buf_hdr;
2208 : : uint64 victim_buf_state;
2209 : 84138342 : uint64 set_bits = 0;
2210 : :
2211 : : /* Make sure we will have room to remember the buffer pin */
2212 : 84138342 : ResourceOwnerEnlarge(CurrentResourceOwner);
2213 : 84138342 : ReservePrivateRefCountEntry();
2214 : :
2215 : : /* create a tag so we can lookup the buffer */
2216 : 84138342 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2217 : :
2218 : : /* determine its hash code and partition lock ID */
2219 : 84138342 : newHash = BufTableHashCode(&newTag);
2220 : 84138342 : newPartitionLock = BufMappingPartitionLock(newHash);
2221 : :
2222 : : /* see if the block is in the buffer pool already */
2223 : 84138342 : LWLockAcquire(newPartitionLock, LW_SHARED);
2224 : 84138342 : existing_buf_id = BufTableLookup(&newTag, newHash);
2225 [ + + ]: 84138342 : if (existing_buf_id >= 0)
2226 : : {
2227 : : BufferDesc *buf;
2228 : : bool valid;
2229 : :
2230 : : /*
2231 : : * Found it. Now, pin the buffer so no one can steal it from the
2232 : : * buffer pool, and check to see if the correct data has been loaded
2233 : : * into the buffer.
2234 : : */
2235 : 82250040 : buf = GetBufferDescriptor(existing_buf_id);
2236 : :
2237 : 82250040 : valid = PinBuffer(buf, strategy, false);
2238 : :
2239 : : /* Can release the mapping lock as soon as we've pinned it */
2240 : 82250040 : LWLockRelease(newPartitionLock);
2241 : :
2242 : 82250040 : *foundPtr = true;
2243 : :
2244 [ + + ]: 82250040 : if (!valid)
2245 : : {
2246 : : /*
2247 : : * We can only get here if (a) someone else is still reading in
2248 : : * the page, (b) a previous read attempt failed, or (c) someone
2249 : : * called StartReadBuffers() but not yet WaitReadBuffers().
2250 : : */
2251 : 2002 : *foundPtr = false;
2252 : : }
2253 : :
2254 : 82250040 : return buf;
2255 : : }
2256 : :
2257 : : /*
2258 : : * Didn't find it in the buffer pool. We'll have to initialize a new
2259 : : * buffer. Remember to unlock the mapping lock while doing the work.
2260 : : */
2261 : 1888302 : LWLockRelease(newPartitionLock);
2262 : :
2263 : : /*
2264 : : * Acquire a victim buffer. Somebody else might try to do the same, we
2265 : : * don't hold any conflicting locks. If so we'll have to undo our work
2266 : : * later.
2267 : : */
2268 : 1888302 : victim_buffer = GetVictimBuffer(strategy, io_context);
2269 : 1888302 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2270 : :
2271 : : /*
2272 : : * Try to make a hashtable entry for the buffer under its new tag. If
2273 : : * somebody else inserted another buffer for the tag, we'll release the
2274 : : * victim buffer we acquired and use the already inserted one.
2275 : : */
2276 : 1888302 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2277 : 1888302 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2278 [ + + ]: 1888302 : if (existing_buf_id >= 0)
2279 : : {
2280 : : BufferDesc *existing_buf_hdr;
2281 : : bool valid;
2282 : :
2283 : : /*
2284 : : * Got a collision. Someone has already done what we were about to do.
2285 : : * We'll just handle this as if it were found in the buffer pool in
2286 : : * the first place. First, give up the buffer we were planning to
2287 : : * use.
2288 : : *
2289 : : * We could do this after releasing the partition lock, but then we'd
2290 : : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2291 : : * before acquiring the lock, for the rare case of such a collision.
2292 : : */
2293 : 612 : UnpinBuffer(victim_buf_hdr);
2294 : :
2295 : : /* remaining code should match code at top of routine */
2296 : :
2297 : 612 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2298 : :
2299 : 612 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2300 : :
2301 : : /* Can release the mapping lock as soon as we've pinned it */
2302 : 612 : LWLockRelease(newPartitionLock);
2303 : :
2304 : 612 : *foundPtr = true;
2305 : :
2306 [ + + ]: 612 : if (!valid)
2307 : : {
2308 : : /*
2309 : : * We can only get here if (a) someone else is still reading in
2310 : : * the page, (b) a previous read attempt failed, or (c) someone
2311 : : * called StartReadBuffers() but not yet WaitReadBuffers().
2312 : : */
2313 : 384 : *foundPtr = false;
2314 : : }
2315 : :
2316 : 612 : return existing_buf_hdr;
2317 : : }
2318 : :
2319 : : /*
2320 : : * Need to lock the buffer header too in order to change its tag.
2321 : : */
2322 : 1887690 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2323 : :
2324 : : /* some sanity checks while we hold the buffer header lock */
2325 : : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2326 : : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2327 : :
2328 : 1887690 : victim_buf_hdr->tag = newTag;
2329 : :
2330 : : /*
2331 : : * Make sure BM_PERMANENT is set for buffers that must be written at every
2332 : : * checkpoint. Unlogged buffers only need to be written at shutdown
2333 : : * checkpoints, except for their "init" forks, which need to be treated
2334 : : * just like permanent relations.
2335 : : */
2336 : 1887690 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2337 [ + + + + ]: 1887690 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2338 : 1887301 : set_bits |= BM_PERMANENT;
2339 : :
2340 : 1887690 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2341 : : set_bits, 0, 0);
2342 : :
2343 : 1887690 : LWLockRelease(newPartitionLock);
2344 : :
2345 : : /*
2346 : : * Buffer contents are currently invalid.
2347 : : */
2348 : 1887690 : *foundPtr = false;
2349 : :
2350 : 1887690 : return victim_buf_hdr;
2351 : : }
2352 : :
2353 : : /*
2354 : : * InvalidateBuffer -- mark a shared buffer invalid.
2355 : : *
2356 : : * The buffer header spinlock must be held at entry. We drop it before
2357 : : * returning. (This is sane because the caller must have locked the
2358 : : * buffer in order to be sure it should be dropped.)
2359 : : *
2360 : : * This is used only in contexts such as dropping a relation. We assume
2361 : : * that no other backend could possibly be interested in using the page,
2362 : : * so the only reason the buffer might be pinned is if someone else is
2363 : : * trying to write it out. We have to let them finish before we can
2364 : : * reclaim the buffer.
2365 : : *
2366 : : * The buffer could get reclaimed by someone else while we are waiting
2367 : : * to acquire the necessary locks; if so, don't mess it up.
2368 : : */
2369 : : static void
2370 : 127443 : InvalidateBuffer(BufferDesc *buf)
2371 : : {
2372 : : BufferTag oldTag;
2373 : : uint32 oldHash; /* hash value for oldTag */
2374 : : LWLock *oldPartitionLock; /* buffer partition lock for it */
2375 : : uint32 oldFlags;
2376 : : uint64 buf_state;
2377 : :
2378 : : /* Save the original buffer tag before dropping the spinlock */
2379 : 127443 : oldTag = buf->tag;
2380 : :
2381 : 127443 : UnlockBufHdr(buf);
2382 : :
2383 : : /*
2384 : : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2385 : : * worth storing the hashcode in BufferDesc so we need not recompute it
2386 : : * here? Probably not.
2387 : : */
2388 : 127443 : oldHash = BufTableHashCode(&oldTag);
2389 : 127443 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2390 : :
2391 : 127443 : retry:
2392 : :
2393 : : /*
2394 : : * Acquire exclusive mapping lock in preparation for changing the buffer's
2395 : : * association.
2396 : : */
2397 : 127443 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2398 : :
2399 : : /* Re-lock the buffer header */
2400 : 127443 : buf_state = LockBufHdr(buf);
2401 : :
2402 : : /* If it's changed while we were waiting for lock, do nothing */
2403 [ - + ]: 127443 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2404 : : {
2405 : 0 : UnlockBufHdr(buf);
2406 : 0 : LWLockRelease(oldPartitionLock);
2407 : 0 : return;
2408 : : }
2409 : :
2410 : : /*
2411 : : * We assume the reason for it to be pinned is that either we were
2412 : : * asynchronously reading the page in before erroring out or someone else
2413 : : * is flushing the page out. Wait for the IO to finish. (This could be
2414 : : * an infinite loop if the refcount is messed up... it would be nice to
2415 : : * time out after awhile, but there seems no way to be sure how many loops
2416 : : * may be needed. Note that if the other guy has pinned the buffer but
2417 : : * not yet done StartBufferIO, WaitIO will fall through and we'll
2418 : : * effectively be busy-looping here.)
2419 : : */
2420 [ - + ]: 127443 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2421 : : {
2422 : 0 : UnlockBufHdr(buf);
2423 : 0 : LWLockRelease(oldPartitionLock);
2424 : : /* safety check: should definitely not be our *own* pin */
2425 [ # # ]: 0 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2426 [ # # ]: 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2427 : 0 : WaitIO(buf);
2428 : 0 : goto retry;
2429 : : }
2430 : :
2431 : : /*
2432 : : * An invalidated buffer should not have any backends waiting to lock the
2433 : : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2434 : : */
2435 : : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2436 : :
2437 : : /*
2438 : : * Clear out the buffer's tag and flags. We must do this to ensure that
2439 : : * linear scans of the buffer array don't think the buffer is valid.
2440 : : */
2441 : 127443 : oldFlags = buf_state & BUF_FLAG_MASK;
2442 : 127443 : ClearBufferTag(&buf->tag);
2443 : :
2444 : 127443 : UnlockBufHdrExt(buf, buf_state,
2445 : : 0,
2446 : : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2447 : : 0);
2448 : :
2449 : : /*
2450 : : * Remove the buffer from the lookup hashtable, if it was in there.
2451 : : */
2452 [ + - ]: 127443 : if (oldFlags & BM_TAG_VALID)
2453 : 127443 : BufTableDelete(&oldTag, oldHash);
2454 : :
2455 : : /*
2456 : : * Done with mapping lock.
2457 : : */
2458 : 127443 : LWLockRelease(oldPartitionLock);
2459 : : }
2460 : :
2461 : : /*
2462 : : * Helper routine for GetVictimBuffer()
2463 : : *
2464 : : * Needs to be called on a buffer with a valid tag, pinned, but without the
2465 : : * buffer header spinlock held.
2466 : : *
2467 : : * Returns true if the buffer can be reused, in which case the buffer is only
2468 : : * pinned by this backend and marked as invalid, false otherwise.
2469 : : */
2470 : : static bool
2471 : 1332857 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2472 : : {
2473 : : uint64 buf_state;
2474 : : uint32 hash;
2475 : : LWLock *partition_lock;
2476 : : BufferTag tag;
2477 : :
2478 : : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2479 : :
2480 : : /* have buffer pinned, so it's safe to read tag without lock */
2481 : 1332857 : tag = buf_hdr->tag;
2482 : :
2483 : 1332857 : hash = BufTableHashCode(&tag);
2484 : 1332857 : partition_lock = BufMappingPartitionLock(hash);
2485 : :
2486 : 1332857 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2487 : :
2488 : : /* lock the buffer header */
2489 : 1332857 : buf_state = LockBufHdr(buf_hdr);
2490 : :
2491 : : /*
2492 : : * We have the buffer pinned nobody else should have been able to unset
2493 : : * this concurrently.
2494 : : */
2495 : : Assert(buf_state & BM_TAG_VALID);
2496 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2497 : : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2498 : :
2499 : : /*
2500 : : * If somebody else pinned the buffer since, or even worse, dirtied it,
2501 : : * give up on this buffer: It's clearly in use.
2502 : : */
2503 [ + + + + ]: 1332857 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2504 : : {
2505 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2506 : :
2507 : 402 : UnlockBufHdr(buf_hdr);
2508 : 402 : LWLockRelease(partition_lock);
2509 : :
2510 : 402 : return false;
2511 : : }
2512 : :
2513 : : /*
2514 : : * An invalidated buffer should not have any backends waiting to lock the
2515 : : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2516 : : */
2517 : : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2518 : :
2519 : : /*
2520 : : * Clear out the buffer's tag and flags and usagecount. This is not
2521 : : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2522 : : * doing anything with the buffer. But currently it's beneficial, as the
2523 : : * cheaper pre-check for several linear scans of shared buffers use the
2524 : : * tag (see e.g. FlushDatabaseBuffers()).
2525 : : */
2526 : 1332455 : ClearBufferTag(&buf_hdr->tag);
2527 : 1332455 : UnlockBufHdrExt(buf_hdr, buf_state,
2528 : : 0,
2529 : : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2530 : : 0);
2531 : :
2532 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2533 : :
2534 : : /* finally delete buffer from the buffer mapping table */
2535 : 1332455 : BufTableDelete(&tag, hash);
2536 : :
2537 : 1332455 : LWLockRelease(partition_lock);
2538 : :
2539 : 1332455 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2540 : : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2541 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2542 : : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u64(&buf_hdr->state)) > 0);
2543 : :
2544 : 1332455 : return true;
2545 : : }
2546 : :
2547 : : static Buffer
2548 : 2170202 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2549 : : {
2550 : : BufferDesc *buf_hdr;
2551 : : Buffer buf;
2552 : : uint64 buf_state;
2553 : : bool from_ring;
2554 : :
2555 : : /*
2556 : : * Ensure, before we pin a victim buffer, that there's a free refcount
2557 : : * entry and resource owner slot for the pin.
2558 : : */
2559 : 2170202 : ReservePrivateRefCountEntry();
2560 : 2170202 : ResourceOwnerEnlarge(CurrentResourceOwner);
2561 : :
2562 : : /* we return here if a prospective victim buffer gets used concurrently */
2563 : 22473 : again:
2564 : :
2565 : : /*
2566 : : * Select a victim buffer. The buffer is returned pinned and owned by
2567 : : * this backend.
2568 : : */
2569 : 2192675 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2570 : 2192675 : buf = BufferDescriptorGetBuffer(buf_hdr);
2571 : :
2572 : : /*
2573 : : * We shouldn't have any other pins for this buffer.
2574 : : */
2575 : 2192675 : CheckBufferIsPinnedOnce(buf);
2576 : :
2577 : : /*
2578 : : * If the buffer was dirty, try to write it out. There is a race
2579 : : * condition here, another backend could dirty the buffer between
2580 : : * StrategyGetBuffer() checking that it is not in use and invalidating the
2581 : : * buffer below. That's addressed by InvalidateVictimBuffer() verifying
2582 : : * that the buffer is not dirty.
2583 : : */
2584 [ + + ]: 2192675 : if (buf_state & BM_DIRTY)
2585 : : {
2586 : : Assert(buf_state & BM_TAG_VALID);
2587 : : Assert(buf_state & BM_VALID);
2588 : :
2589 : : /*
2590 : : * We need a share-exclusive lock on the buffer contents to write it
2591 : : * out (else we might write invalid data, eg because someone else is
2592 : : * compacting the page contents while we write). We must use a
2593 : : * conditional lock acquisition here to avoid deadlock. Even though
2594 : : * the buffer was not pinned (and therefore surely not locked) when
2595 : : * StrategyGetBuffer returned it, someone else could have pinned and
2596 : : * (share-)exclusive-locked it by the time we get here. If we try to
2597 : : * get the lock unconditionally, we'd block waiting for them; if they
2598 : : * later block waiting for us, deadlock ensues. (This has been
2599 : : * observed to happen when two backends are both trying to split btree
2600 : : * index pages, and the second one just happens to be trying to split
2601 : : * the page the first one got from StrategyGetBuffer.)
2602 : : */
2603 [ - + ]: 378031 : if (!BufferLockConditional(buf, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE))
2604 : : {
2605 : : /*
2606 : : * Someone else has locked the buffer, so give it up and loop back
2607 : : * to get another one.
2608 : : */
2609 : 0 : UnpinBuffer(buf_hdr);
2610 : 0 : goto again;
2611 : : }
2612 : :
2613 : : /*
2614 : : * If using a nondefault strategy, and this victim came from the
2615 : : * strategy ring, let the strategy decide whether to reject it when
2616 : : * reusing it would require a WAL flush. This only applies to
2617 : : * permanent buffers; unlogged buffers can have fake LSNs, so
2618 : : * XLogNeedsFlush() is not meaningful for them.
2619 : : *
2620 : : * We need to hold the content lock in at least share-exclusive mode
2621 : : * to safely inspect the page LSN, so this couldn't have been done
2622 : : * inside StrategyGetBuffer().
2623 : : */
2624 [ + + + + ]: 378031 : if (strategy && from_ring &&
2625 [ + + + + ]: 196111 : buf_state & BM_PERMANENT &&
2626 [ + + ]: 124738 : XLogNeedsFlush(BufferGetLSN(buf_hdr)) &&
2627 : 26692 : StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2628 : : {
2629 : 22071 : UnlockReleaseBuffer(buf);
2630 : 22071 : goto again;
2631 : : }
2632 : :
2633 : : /* OK, do the I/O */
2634 : 355960 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2635 : 355960 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
2636 : :
2637 : 355960 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2638 : : &buf_hdr->tag);
2639 : : }
2640 : :
2641 : :
2642 [ + + ]: 2170604 : if (buf_state & BM_VALID)
2643 : : {
2644 : : /*
2645 : : * When a BufferAccessStrategy is in use, blocks evicted from shared
2646 : : * buffers are counted as IOOP_EVICT in the corresponding context
2647 : : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2648 : : * strategy in two cases: 1) while initially claiming buffers for the
2649 : : * strategy ring 2) to replace an existing strategy ring buffer
2650 : : * because it is pinned or in use and cannot be reused.
2651 : : *
2652 : : * Blocks evicted from buffers already in the strategy ring are
2653 : : * counted as IOOP_REUSE in the corresponding strategy context.
2654 : : *
2655 : : * At this point, we can accurately count evictions and reuses,
2656 : : * because we have successfully claimed the valid buffer. Previously,
2657 : : * we may have been forced to release the buffer due to concurrent
2658 : : * pinners or erroring out.
2659 : : */
2660 : 1330339 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2661 [ + + ]: 1330339 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2662 : : }
2663 : :
2664 : : /*
2665 : : * If the buffer has an entry in the buffer mapping table, delete it. This
2666 : : * can fail because another backend could have pinned or dirtied the
2667 : : * buffer.
2668 : : */
2669 [ + + + + ]: 2170604 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2670 : : {
2671 : 402 : UnpinBuffer(buf_hdr);
2672 : 402 : goto again;
2673 : : }
2674 : :
2675 : : /* a final set of sanity checks */
2676 : : #ifdef USE_ASSERT_CHECKING
2677 : : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2678 : :
2679 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2680 : : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2681 : :
2682 : : CheckBufferIsPinnedOnce(buf);
2683 : : #endif
2684 : :
2685 : 2170202 : return buf;
2686 : : }
2687 : :
2688 : : /*
2689 : : * Return the maximum number of buffers that a backend should try to pin once,
2690 : : * to avoid exceeding its fair share. This is the highest value that
2691 : : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2692 : : * system with a very small buffer pool relative to max_connections.
2693 : : */
2694 : : uint32
2695 : 765890 : GetPinLimit(void)
2696 : : {
2697 : 765890 : return MaxProportionalPins;
2698 : : }
2699 : :
2700 : : /*
2701 : : * Return the maximum number of additional buffers that this backend should
2702 : : * pin if it wants to stay under the per-backend limit, considering the number
2703 : : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2704 : : * return by this function can be zero.
2705 : : */
2706 : : uint32
2707 : 4199438 : GetAdditionalPinLimit(void)
2708 : : {
2709 : : uint32 estimated_pins_held;
2710 : :
2711 : : /*
2712 : : * We get the number of "overflowed" pins for free, but don't know the
2713 : : * number of pins in PrivateRefCountArray. The cost of calculating that
2714 : : * exactly doesn't seem worth it, so just assume the max.
2715 : : */
2716 : 4199438 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2717 : :
2718 : : /* Is this backend already holding more than its fair share? */
2719 [ + + ]: 4199438 : if (estimated_pins_held > MaxProportionalPins)
2720 : 1341112 : return 0;
2721 : :
2722 : 2858326 : return MaxProportionalPins - estimated_pins_held;
2723 : : }
2724 : :
2725 : : /*
2726 : : * Limit the number of pins a batch operation may additionally acquire, to
2727 : : * avoid running out of pinnable buffers.
2728 : : *
2729 : : * One additional pin is always allowed, on the assumption that the operation
2730 : : * requires at least one to make progress.
2731 : : */
2732 : : void
2733 : 254998 : LimitAdditionalPins(uint32 *additional_pins)
2734 : : {
2735 : : uint32 limit;
2736 : :
2737 [ + + ]: 254998 : if (*additional_pins <= 1)
2738 : 239380 : return;
2739 : :
2740 : 15618 : limit = GetAdditionalPinLimit();
2741 : 15618 : limit = Max(limit, 1);
2742 [ + + ]: 15618 : if (limit < *additional_pins)
2743 : 10068 : *additional_pins = limit;
2744 : : }
2745 : :
2746 : : /*
2747 : : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2748 : : * avoid duplicating the tracing and relpersistence related logic.
2749 : : */
2750 : : static BlockNumber
2751 : 269968 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2752 : : ForkNumber fork,
2753 : : BufferAccessStrategy strategy,
2754 : : uint32 flags,
2755 : : uint32 extend_by,
2756 : : BlockNumber extend_upto,
2757 : : Buffer *buffers,
2758 : : uint32 *extended_by)
2759 : : {
2760 : : BlockNumber first_block;
2761 : :
2762 : : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2763 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2764 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2765 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2766 : : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2767 : : extend_by);
2768 : :
2769 [ + + ]: 269968 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2770 : : {
2771 : : /*
2772 : : * Reject attempts to extend non-local temporary relations; we have no
2773 : : * ability to transfer about-to-be-created local buffers into the
2774 : : * owning session's local buffers. This is the canonical place for
2775 : : * the check, covering any attempt to extend a non-local temporary
2776 : : * relation.
2777 : : */
2778 [ + - + - : 14970 : if (bmr.rel && RELATION_IS_OTHER_TEMP(bmr.rel))
+ + ]
2779 [ + - ]: 1 : ereport(ERROR,
2780 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2781 : : errmsg("cannot access temporary tables of other sessions")));
2782 : :
2783 : 14969 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2784 : : extend_by, extend_upto,
2785 : : buffers, &extend_by);
2786 : : }
2787 : : else
2788 : 254998 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2789 : : extend_by, extend_upto,
2790 : : buffers, &extend_by);
2791 : 269967 : *extended_by = extend_by;
2792 : :
2793 : : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2794 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2795 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2796 : : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2797 : : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2798 : : *extended_by,
2799 : : first_block);
2800 : :
2801 : 269967 : return first_block;
2802 : : }
2803 : :
2804 : : /*
2805 : : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2806 : : * shared buffers.
2807 : : */
2808 : : static BlockNumber
2809 : 254998 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2810 : : ForkNumber fork,
2811 : : BufferAccessStrategy strategy,
2812 : : uint32 flags,
2813 : : uint32 extend_by,
2814 : : BlockNumber extend_upto,
2815 : : Buffer *buffers,
2816 : : uint32 *extended_by)
2817 : : {
2818 : : BlockNumber first_block;
2819 : 254998 : IOContext io_context = IOContextForStrategy(strategy);
2820 : : instr_time io_start;
2821 : :
2822 : 254998 : LimitAdditionalPins(&extend_by);
2823 : :
2824 : : /*
2825 : : * Acquire victim buffers for extension without holding extension lock.
2826 : : * Writing out victim buffers is the most expensive part of extending the
2827 : : * relation, particularly when doing so requires WAL flushes. Zeroing out
2828 : : * the buffers is also quite expensive, so do that before holding the
2829 : : * extension lock as well.
2830 : : *
2831 : : * These pages are pinned by us and not valid. While we hold the pin they
2832 : : * can't be acquired as victim buffers by another backend.
2833 : : */
2834 [ + + ]: 536898 : for (uint32 i = 0; i < extend_by; i++)
2835 : : {
2836 : : Block buf_block;
2837 : :
2838 : 281900 : buffers[i] = GetVictimBuffer(strategy, io_context);
2839 : 281900 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2840 : :
2841 : : /* new buffers are zero-filled */
2842 [ + - + - : 281900 : MemSet(buf_block, 0, BLCKSZ);
+ - - + -
- ]
2843 : : }
2844 : :
2845 : : /*
2846 : : * Lock relation against concurrent extensions, unless requested not to.
2847 : : *
2848 : : * We use the same extension lock for all forks. That's unnecessarily
2849 : : * restrictive, but currently extensions for forks don't happen often
2850 : : * enough to make it worth locking more granularly.
2851 : : *
2852 : : * Note that another backend might have extended the relation by the time
2853 : : * we get the lock.
2854 : : */
2855 [ + + ]: 254998 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2856 : 199290 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2857 : :
2858 : : /*
2859 : : * If requested, invalidate size cache, so that smgrnblocks asks the
2860 : : * kernel.
2861 : : */
2862 [ + + ]: 254998 : if (flags & EB_CLEAR_SIZE_CACHE)
2863 [ + - ]: 10035 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2864 : :
2865 [ + + ]: 254998 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2866 : :
2867 : : /*
2868 : : * Now that we have the accurate relation size, check if the caller wants
2869 : : * us to extend to only up to a specific size. If there were concurrent
2870 : : * extensions, we might have acquired too many buffers and need to release
2871 : : * them.
2872 : : */
2873 [ + + ]: 254998 : if (extend_upto != InvalidBlockNumber)
2874 : : {
2875 : 57899 : uint32 orig_extend_by = extend_by;
2876 : :
2877 [ - + ]: 57899 : if (first_block > extend_upto)
2878 : 0 : extend_by = 0;
2879 [ + + ]: 57899 : else if ((uint64) first_block + extend_by > extend_upto)
2880 : 6 : extend_by = extend_upto - first_block;
2881 : :
2882 [ + + ]: 57909 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2883 : : {
2884 : 10 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2885 : :
2886 : 10 : UnpinBuffer(buf_hdr);
2887 : : }
2888 : :
2889 [ + + ]: 57899 : if (extend_by == 0)
2890 : : {
2891 [ + - ]: 6 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2892 : 6 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2893 : 6 : *extended_by = extend_by;
2894 : 6 : return first_block;
2895 : : }
2896 : : }
2897 : :
2898 : : /* Fail if relation is already at maximum possible length */
2899 [ - + ]: 254992 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2900 [ # # # # : 0 : ereport(ERROR,
# # # # #
# ]
2901 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2902 : : errmsg("cannot extend relation %s beyond %u blocks",
2903 : : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2904 : : MaxBlockNumber)));
2905 : :
2906 : : /*
2907 : : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2908 : : *
2909 : : * This needs to happen before we extend the relation, because as soon as
2910 : : * we do, other backends can start to read in those pages.
2911 : : */
2912 [ + + ]: 536882 : for (uint32 i = 0; i < extend_by; i++)
2913 : : {
2914 : 281890 : Buffer victim_buf = buffers[i];
2915 : 281890 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2916 : : BufferTag tag;
2917 : : uint32 hash;
2918 : : LWLock *partition_lock;
2919 : : int existing_id;
2920 : :
2921 : : /* in case we need to pin an existing buffer below */
2922 : 281890 : ResourceOwnerEnlarge(CurrentResourceOwner);
2923 : 281890 : ReservePrivateRefCountEntry();
2924 : :
2925 [ + + ]: 281890 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2926 : : first_block + i);
2927 : 281890 : hash = BufTableHashCode(&tag);
2928 : 281890 : partition_lock = BufMappingPartitionLock(hash);
2929 : :
2930 : 281890 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2931 : :
2932 : 281890 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2933 : :
2934 : : /*
2935 : : * We get here only in the corner case where we are trying to extend
2936 : : * the relation but we found a pre-existing buffer. This can happen
2937 : : * because a prior attempt at extending the relation failed, and
2938 : : * because mdread doesn't complain about reads beyond EOF (when
2939 : : * zero_damaged_pages is ON) and so a previous attempt to read a block
2940 : : * beyond EOF could have left a "valid" zero-filled buffer.
2941 : : *
2942 : : * This has also been observed when relation was overwritten by
2943 : : * external process. Since the legitimate cases should always have
2944 : : * left a zero-filled buffer, complain if not PageIsNew.
2945 : : */
2946 [ - + ]: 281890 : if (existing_id >= 0)
2947 : : {
2948 : 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2949 : : Block buf_block;
2950 : : bool valid;
2951 : :
2952 : : /*
2953 : : * Pin the existing buffer before releasing the partition lock,
2954 : : * preventing it from being evicted.
2955 : : */
2956 : 0 : valid = PinBuffer(existing_hdr, strategy, false);
2957 : :
2958 : 0 : LWLockRelease(partition_lock);
2959 : 0 : UnpinBuffer(victim_buf_hdr);
2960 : :
2961 : 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2962 : 0 : buf_block = BufHdrGetBlock(existing_hdr);
2963 : :
2964 [ # # # # ]: 0 : if (valid && !PageIsNew((Page) buf_block))
2965 [ # # # # : 0 : ereport(ERROR,
# # # # #
# ]
2966 : : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2967 : : existing_hdr->tag.blockNum,
2968 : : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2969 : :
2970 : : /*
2971 : : * We *must* do smgr[zero]extend before succeeding, else the page
2972 : : * will not be reserved by the kernel, and the next P_NEW call
2973 : : * will decide to return the same page. Clear the BM_VALID bit,
2974 : : * do StartSharedBufferIO() and proceed.
2975 : : *
2976 : : * Loop to handle the very small possibility that someone re-sets
2977 : : * BM_VALID between our clearing it and StartSharedBufferIO
2978 : : * inspecting it.
2979 : : */
2980 : : while (true)
2981 : 0 : {
2982 : : StartBufferIOResult sbres;
2983 : :
2984 : 0 : pg_atomic_fetch_and_u64(&existing_hdr->state, ~BM_VALID);
2985 : :
2986 : 0 : sbres = StartSharedBufferIO(existing_hdr, true, true, NULL);
2987 : :
2988 [ # # ]: 0 : if (sbres != BUFFER_IO_ALREADY_DONE)
2989 : 0 : break;
2990 : : }
2991 : : }
2992 : : else
2993 : : {
2994 : : uint64 buf_state;
2995 : 281890 : uint64 set_bits = 0;
2996 : :
2997 : 281890 : buf_state = LockBufHdr(victim_buf_hdr);
2998 : :
2999 : : /* some sanity checks while we hold the buffer header lock */
3000 : : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY)));
3001 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
3002 : :
3003 : 281890 : victim_buf_hdr->tag = tag;
3004 : :
3005 : 281890 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
3006 [ + + + + ]: 281890 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
3007 : 276462 : set_bits |= BM_PERMANENT;
3008 : :
3009 : 281890 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
3010 : : set_bits, 0,
3011 : : 0);
3012 : :
3013 : 281890 : LWLockRelease(partition_lock);
3014 : :
3015 : : /* XXX: could combine the locked operations in it with the above */
3016 : 281890 : StartSharedBufferIO(victim_buf_hdr, true, true, NULL);
3017 : : }
3018 : : }
3019 : :
3020 : 254992 : io_start = pgstat_prepare_io_time(track_io_timing);
3021 : :
3022 : : /*
3023 : : * Note: if smgrzeroextend fails, we will end up with buffers that are
3024 : : * allocated but not marked BM_VALID. The next relation extension will
3025 : : * still select the same block number (because the relation didn't get any
3026 : : * longer on disk) and so future attempts to extend the relation will find
3027 : : * the same buffers (if they have not been recycled) but come right back
3028 : : * here to try smgrzeroextend again.
3029 : : *
3030 : : * We don't need to set checksum for all-zero pages.
3031 : : */
3032 [ + + ]: 254992 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
3033 : :
3034 : : /*
3035 : : * Release the file-extension lock; it's now OK for someone else to extend
3036 : : * the relation some more.
3037 : : *
3038 : : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
3039 : : * take noticeable time.
3040 : : */
3041 [ + + ]: 254992 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
3042 : 199284 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
3043 : :
3044 : 254992 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
3045 : 254992 : io_start, 1, extend_by * BLCKSZ);
3046 : :
3047 : : /* Set BM_VALID, terminate IO, and wake up any waiters */
3048 [ + + ]: 536882 : for (uint32 i = 0; i < extend_by; i++)
3049 : : {
3050 : 281890 : Buffer buf = buffers[i];
3051 : 281890 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
3052 : 281890 : bool lock = false;
3053 : :
3054 [ + + + + ]: 281890 : if (flags & EB_LOCK_FIRST && i == 0)
3055 : 196771 : lock = true;
3056 [ + + ]: 85119 : else if (flags & EB_LOCK_TARGET)
3057 : : {
3058 : : Assert(extend_upto != InvalidBlockNumber);
3059 [ + + ]: 46712 : if (first_block + i + 1 == extend_upto)
3060 : 46140 : lock = true;
3061 : : }
3062 : :
3063 [ + + ]: 281890 : if (lock)
3064 : 242911 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
3065 : :
3066 : 281890 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
3067 : : }
3068 : :
3069 : 254992 : pgBufferUsage.shared_blks_written += extend_by;
3070 : :
3071 : 254992 : *extended_by = extend_by;
3072 : :
3073 : 254992 : return first_block;
3074 : : }
3075 : :
3076 : : /*
3077 : : * BufferIsLockedByMe
3078 : : *
3079 : : * Checks if this backend has the buffer locked in any mode.
3080 : : *
3081 : : * Buffer must be pinned.
3082 : : */
3083 : : bool
3084 : 0 : BufferIsLockedByMe(Buffer buffer)
3085 : : {
3086 : : BufferDesc *bufHdr;
3087 : :
3088 : : Assert(BufferIsPinned(buffer));
3089 : :
3090 [ # # ]: 0 : if (BufferIsLocal(buffer))
3091 : : {
3092 : : /* Content locks are not maintained for local buffers. */
3093 : 0 : return true;
3094 : : }
3095 : : else
3096 : : {
3097 : 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3098 : 0 : return BufferLockHeldByMe(bufHdr);
3099 : : }
3100 : : }
3101 : :
3102 : : /*
3103 : : * BufferIsLockedByMeInMode
3104 : : *
3105 : : * Checks if this backend has the buffer locked in the specified mode.
3106 : : *
3107 : : * Buffer must be pinned.
3108 : : */
3109 : : bool
3110 : 0 : BufferIsLockedByMeInMode(Buffer buffer, BufferLockMode mode)
3111 : : {
3112 : : BufferDesc *bufHdr;
3113 : :
3114 : : Assert(BufferIsPinned(buffer));
3115 : :
3116 [ # # ]: 0 : if (BufferIsLocal(buffer))
3117 : : {
3118 : : /* Content locks are not maintained for local buffers. */
3119 : 0 : return true;
3120 : : }
3121 : : else
3122 : : {
3123 : 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3124 : 0 : return BufferLockHeldByMeInMode(bufHdr, mode);
3125 : : }
3126 : : }
3127 : :
3128 : : /*
3129 : : * BufferIsDirty
3130 : : *
3131 : : * Checks if buffer is already dirty.
3132 : : *
3133 : : * Buffer must be pinned and [share-]exclusive-locked. (Without such a lock,
3134 : : * the result may be stale before it's returned.)
3135 : : */
3136 : : bool
3137 : 29470 : BufferIsDirty(Buffer buffer)
3138 : : {
3139 : : BufferDesc *bufHdr;
3140 : :
3141 : : Assert(BufferIsPinned(buffer));
3142 : :
3143 [ + + ]: 29470 : if (BufferIsLocal(buffer))
3144 : : {
3145 : 9523 : int bufid = -buffer - 1;
3146 : :
3147 : 9523 : bufHdr = GetLocalBufferDescriptor(bufid);
3148 : : /* Content locks are not maintained for local buffers. */
3149 : : }
3150 : : else
3151 : : {
3152 : 19947 : bufHdr = GetBufferDescriptor(buffer - 1);
3153 : : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
3154 : : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3155 : : }
3156 : :
3157 : 29470 : return pg_atomic_read_u64(&bufHdr->state) & BM_DIRTY;
3158 : : }
3159 : :
3160 : : /*
3161 : : * MarkBufferDirty
3162 : : *
3163 : : * Marks buffer contents as dirty (actual write happens later).
3164 : : *
3165 : : * Buffer must be pinned and exclusive-locked. (If caller does not hold
3166 : : * exclusive lock, then somebody could be in process of writing the buffer,
3167 : : * leading to risk of bad data written to disk.)
3168 : : */
3169 : : void
3170 : 34868609 : MarkBufferDirty(Buffer buffer)
3171 : : {
3172 : : BufferDesc *bufHdr;
3173 : : uint64 buf_state;
3174 : : uint64 old_buf_state;
3175 : :
3176 [ - + ]: 34868609 : if (!BufferIsValid(buffer))
3177 [ # # ]: 0 : elog(ERROR, "bad buffer ID: %d", buffer);
3178 : :
3179 [ + + ]: 34868609 : if (BufferIsLocal(buffer))
3180 : : {
3181 : 1580070 : MarkLocalBufferDirty(buffer);
3182 : 1580070 : return;
3183 : : }
3184 : :
3185 : 33288539 : bufHdr = GetBufferDescriptor(buffer - 1);
3186 : :
3187 : : Assert(BufferIsPinned(buffer));
3188 : : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3189 : :
3190 : : /*
3191 : : * NB: We have to wait for the buffer header spinlock to be not held, as
3192 : : * TerminateBufferIO() relies on the spinlock.
3193 : : */
3194 : 33288539 : old_buf_state = pg_atomic_read_u64(&bufHdr->state);
3195 : : for (;;)
3196 : : {
3197 [ + + ]: 33288870 : if (old_buf_state & BM_LOCKED)
3198 : 489 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
3199 : :
3200 : 33288870 : buf_state = old_buf_state;
3201 : :
3202 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
3203 : 33288870 : buf_state |= BM_DIRTY;
3204 : :
3205 [ + + ]: 33288870 : if (pg_atomic_compare_exchange_u64(&bufHdr->state, &old_buf_state,
3206 : : buf_state))
3207 : 33288539 : break;
3208 : : }
3209 : :
3210 : : /*
3211 : : * If the buffer was not dirty already, do vacuum accounting.
3212 : : */
3213 [ + + ]: 33288539 : if (!(old_buf_state & BM_DIRTY))
3214 : : {
3215 : 800673 : pgBufferUsage.shared_blks_dirtied++;
3216 [ + + ]: 800673 : if (VacuumCostActive)
3217 : 8508 : VacuumCostBalance += VacuumCostPageDirty;
3218 : : }
3219 : : }
3220 : :
3221 : : /*
3222 : : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
3223 : : *
3224 : : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
3225 : : * compared to calling the two routines separately. Now it's mainly just
3226 : : * a convenience function. However, if the passed buffer is valid and
3227 : : * already contains the desired block, we just return it as-is; and that
3228 : : * does save considerable work compared to a full release and reacquire.
3229 : : *
3230 : : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3231 : : * buffer actually needs to be released. This case is the same as ReadBuffer,
3232 : : * but can save some tests in the caller.
3233 : : */
3234 : : Buffer
3235 : 3646 : ReleaseAndReadBuffer(Buffer buffer,
3236 : : Relation relation,
3237 : : BlockNumber blockNum)
3238 : : {
3239 : 3646 : ForkNumber forkNum = MAIN_FORKNUM;
3240 : : BufferDesc *bufHdr;
3241 : :
3242 [ + - ]: 3646 : if (BufferIsValid(buffer))
3243 : : {
3244 : : Assert(BufferIsPinned(buffer));
3245 [ + + ]: 3646 : if (BufferIsLocal(buffer))
3246 : : {
3247 : 50 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3248 [ - + - - ]: 50 : if (bufHdr->tag.blockNum == blockNum &&
3249 [ # # ]: 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3250 : 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3251 : 0 : return buffer;
3252 : 50 : UnpinLocalBuffer(buffer);
3253 : : }
3254 : : else
3255 : : {
3256 : 3596 : bufHdr = GetBufferDescriptor(buffer - 1);
3257 : : /* we have pin, so it's ok to examine tag without spinlock */
3258 [ - + - - ]: 3596 : if (bufHdr->tag.blockNum == blockNum &&
3259 [ # # ]: 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3260 : 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3261 : 0 : return buffer;
3262 : 3596 : UnpinBuffer(bufHdr);
3263 : : }
3264 : : }
3265 : :
3266 : 3646 : return ReadBuffer(relation, blockNum);
3267 : : }
3268 : :
3269 : : /*
3270 : : * PinBuffer -- make buffer unavailable for replacement.
3271 : : *
3272 : : * For the default access strategy, the buffer's usage_count is incremented
3273 : : * when we first pin it; for other strategies we just make sure the usage_count
3274 : : * isn't zero. (The idea of the latter is that we don't want synchronized
3275 : : * heap scans to inflate the count, but we need it to not be zero to discourage
3276 : : * other backends from stealing buffers from our ring. As long as we cycle
3277 : : * through the ring faster than the global clock-sweep cycles, buffers in
3278 : : * our ring won't be chosen as victims for replacement by other backends.)
3279 : : *
3280 : : * This should be applied only to shared buffers, never local ones.
3281 : : *
3282 : : * Since buffers are pinned/unpinned very frequently, pin buffers without
3283 : : * taking the buffer header lock; instead update the state variable in loop of
3284 : : * CAS operations. Hopefully it's just a single CAS.
3285 : : *
3286 : : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3287 : : * must have been done already.
3288 : : *
3289 : : * Returns true if buffer is BM_VALID, else false. This provision allows
3290 : : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3291 : : * true, then a false return value also indicates that the buffer was
3292 : : * (recently) invalid and has not been pinned.
3293 : : */
3294 : : static bool
3295 : 82255593 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3296 : : bool skip_if_not_valid)
3297 : : {
3298 : 82255593 : Buffer b = BufferDescriptorGetBuffer(buf);
3299 : : bool result;
3300 : : PrivateRefCountEntry *ref;
3301 : :
3302 : : Assert(!BufferIsLocal(b));
3303 : : Assert(ReservedRefCountSlot != -1);
3304 : :
3305 : 82255593 : ref = GetPrivateRefCountEntry(b, true);
3306 : :
3307 [ + + ]: 82255593 : if (ref == NULL)
3308 : : {
3309 : : uint64 buf_state;
3310 : : uint64 old_buf_state;
3311 : :
3312 : 74819155 : old_buf_state = pg_atomic_read_u64(&buf->state);
3313 : : for (;;)
3314 : : {
3315 [ + + + + : 74842936 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
+ + ]
3316 : 6 : return false;
3317 : :
3318 : : /*
3319 : : * We're not allowed to increase the refcount while the buffer
3320 : : * header spinlock is held. Wait for the lock to be released.
3321 : : */
3322 [ + + ]: 74842930 : if (unlikely(old_buf_state & BM_LOCKED))
3323 : : {
3324 : 157 : old_buf_state = WaitBufHdrUnlocked(buf);
3325 : :
3326 : : /* perform checks at the top of the loop again */
3327 : 157 : continue;
3328 : : }
3329 : :
3330 : 74842773 : buf_state = old_buf_state;
3331 : :
3332 : : /* increase refcount */
3333 : 74842773 : buf_state += BUF_REFCOUNT_ONE;
3334 : :
3335 [ + + ]: 74842773 : if (strategy == NULL)
3336 : : {
3337 : : /* Default case: increase usagecount unless already max. */
3338 [ + + ]: 74010288 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3339 : 3781941 : buf_state += BUF_USAGECOUNT_ONE;
3340 : : }
3341 : : else
3342 : : {
3343 : : /*
3344 : : * Ring buffers shouldn't evict others from pool. Thus we
3345 : : * don't make usagecount more than 1.
3346 : : */
3347 [ + + ]: 832485 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3348 : 33917 : buf_state += BUF_USAGECOUNT_ONE;
3349 : : }
3350 : :
3351 [ + + ]: 74842773 : if (pg_atomic_compare_exchange_u64(&buf->state, &old_buf_state,
3352 : : buf_state))
3353 : : {
3354 : 74819149 : result = (buf_state & BM_VALID) != 0;
3355 : :
3356 : 74819149 : TrackNewBufferPin(b);
3357 : 74819149 : break;
3358 : : }
3359 : : }
3360 : : }
3361 : : else
3362 : : {
3363 : : /*
3364 : : * If we previously pinned the buffer, it is likely to be valid, but
3365 : : * it may not be if StartReadBuffers() was called and
3366 : : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3367 : : * the flags without locking. This is racy, but it's OK to return
3368 : : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3369 : : * it'll see that it's now valid.
3370 : : *
3371 : : * Note: We deliberately avoid a Valgrind client request here.
3372 : : * Individual access methods can optionally superimpose buffer page
3373 : : * client requests on top of our client requests to enforce that
3374 : : * buffers are only accessed while locked (and pinned). It's possible
3375 : : * that the buffer page is legitimately non-accessible here. We
3376 : : * cannot meddle with that.
3377 : : */
3378 : 7436438 : result = (pg_atomic_read_u64(&buf->state) & BM_VALID) != 0;
3379 : :
3380 : : Assert(ref->data.refcount > 0);
3381 : 7436438 : ref->data.refcount++;
3382 : 7436438 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3383 : : }
3384 : :
3385 : 82255587 : return result;
3386 : : }
3387 : :
3388 : : /*
3389 : : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3390 : : * The spinlock is released before return.
3391 : : *
3392 : : * As this function is called with the spinlock held, the caller has to
3393 : : * previously call ReservePrivateRefCountEntry() and
3394 : : * ResourceOwnerEnlarge(CurrentResourceOwner);
3395 : : *
3396 : : * Currently, no callers of this function want to modify the buffer's
3397 : : * usage_count at all, so there's no need for a strategy parameter.
3398 : : * Also we don't bother with a BM_VALID test (the caller could check that for
3399 : : * itself).
3400 : : *
3401 : : * Also all callers only ever use this function when it's known that the
3402 : : * buffer can't have a preexisting pin by this backend. That allows us to skip
3403 : : * searching the private refcount array & hash, which is a boon, because the
3404 : : * spinlock is still held.
3405 : : *
3406 : : * Note: use of this routine is frequently mandatory, not just an optimization
3407 : : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3408 : : * its state can change under us.
3409 : : */
3410 : : static void
3411 : 345815 : PinBuffer_Locked(BufferDesc *buf)
3412 : : {
3413 : : uint64 old_buf_state;
3414 : :
3415 : : /*
3416 : : * As explained, We don't expect any preexisting pins. That allows us to
3417 : : * manipulate the PrivateRefCount after releasing the spinlock
3418 : : */
3419 : : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3420 : :
3421 : : /*
3422 : : * Since we hold the buffer spinlock, we can update the buffer state and
3423 : : * release the lock in one operation.
3424 : : */
3425 : 345815 : old_buf_state = pg_atomic_read_u64(&buf->state);
3426 : :
3427 : 345815 : UnlockBufHdrExt(buf, old_buf_state,
3428 : : 0, 0, 1);
3429 : :
3430 : 345815 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3431 : 345815 : }
3432 : :
3433 : : /*
3434 : : * Support for waking up another backend that is waiting for the cleanup lock
3435 : : * to be released using BM_PIN_COUNT_WAITER.
3436 : : *
3437 : : * See LockBufferForCleanup().
3438 : : *
3439 : : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3440 : : * not just reducing the backend-local pincount for the buffer).
3441 : : */
3442 : : static void
3443 : 5 : WakePinCountWaiter(BufferDesc *buf)
3444 : : {
3445 : : /*
3446 : : * Acquire the buffer header lock, re-check that there's a waiter. Another
3447 : : * backend could have unpinned this buffer, and already woken up the
3448 : : * waiter.
3449 : : *
3450 : : * There's no danger of the buffer being replaced after we unpinned it
3451 : : * above, as it's pinned by the waiter. The waiter removes
3452 : : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3453 : : * backend waking it up.
3454 : : */
3455 : 5 : uint64 buf_state = LockBufHdr(buf);
3456 : :
3457 [ + - ]: 5 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3458 [ + - ]: 5 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3459 : 5 : {
3460 : : /* we just released the last pin other than the waiter's */
3461 : 5 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3462 : :
3463 : 5 : UnlockBufHdrExt(buf, buf_state,
3464 : : 0, BM_PIN_COUNT_WAITER,
3465 : : 0);
3466 : 5 : ProcSendSignal(wait_backend_pgprocno);
3467 : : }
3468 : : else
3469 : 0 : UnlockBufHdr(buf);
3470 : 5 : }
3471 : :
3472 : : /*
3473 : : * UnpinBuffer -- make buffer available for replacement.
3474 : : *
3475 : : * This should be applied only to shared buffers, never local ones. This
3476 : : * always adjusts CurrentResourceOwner.
3477 : : */
3478 : : static void
3479 : 49624078 : UnpinBuffer(BufferDesc *buf)
3480 : : {
3481 : 49624078 : Buffer b = BufferDescriptorGetBuffer(buf);
3482 : :
3483 : 49624078 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3484 : 49624078 : UnpinBufferNoOwner(buf);
3485 : 49624078 : }
3486 : :
3487 : : static void
3488 : 49630815 : UnpinBufferNoOwner(BufferDesc *buf)
3489 : : {
3490 : : PrivateRefCountEntry *ref;
3491 : 49630815 : Buffer b = BufferDescriptorGetBuffer(buf);
3492 : :
3493 : : Assert(!BufferIsLocal(b));
3494 : :
3495 : : /* not moving as we're likely deleting it soon anyway */
3496 : 49630815 : ref = GetPrivateRefCountEntry(b, false);
3497 : : Assert(ref != NULL);
3498 : : Assert(ref->data.refcount > 0);
3499 : 49630815 : ref->data.refcount--;
3500 [ + + ]: 49630815 : if (ref->data.refcount == 0)
3501 : : {
3502 : : uint64 old_buf_state;
3503 : :
3504 : : /*
3505 : : * Mark buffer non-accessible to Valgrind.
3506 : : *
3507 : : * Note that the buffer may have already been marked non-accessible
3508 : : * within access method code that enforces that buffers are only
3509 : : * accessed while a buffer lock is held.
3510 : : */
3511 : : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3512 : :
3513 : : /*
3514 : : * I'd better not still hold the buffer content lock. Can't use
3515 : : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3516 : : */
3517 : : Assert(!BufferLockHeldByMe(buf));
3518 : :
3519 : : /* decrement the shared reference count */
3520 : 30001665 : old_buf_state = pg_atomic_fetch_sub_u64(&buf->state, BUF_REFCOUNT_ONE);
3521 : :
3522 : : /* Support LockBufferForCleanup() */
3523 [ + + ]: 30001665 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3524 : 4 : WakePinCountWaiter(buf);
3525 : :
3526 : 30001665 : ForgetPrivateRefCountEntry(ref);
3527 : : }
3528 : 49630815 : }
3529 : :
3530 : : /*
3531 : : * Set up backend-local tracking of a buffer pinned the first time by this
3532 : : * backend.
3533 : : */
3534 : : inline void
3535 : 77357639 : TrackNewBufferPin(Buffer buf)
3536 : : {
3537 : : PrivateRefCountEntry *ref;
3538 : :
3539 : 77357639 : ref = NewPrivateRefCountEntry(buf);
3540 : 77357639 : ref->data.refcount++;
3541 : :
3542 : 77357639 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3543 : :
3544 : : /*
3545 : : * This is the first pin for this page by this backend, mark its page as
3546 : : * defined to valgrind. While the page contents might not actually be
3547 : : * valid yet, we don't currently guarantee that such pages are marked
3548 : : * undefined or non-accessible.
3549 : : *
3550 : : * It's not necessarily the prettiest to do this here, but otherwise we'd
3551 : : * need this block of code in multiple places.
3552 : : */
3553 : : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3554 : : BLCKSZ);
3555 : 77357639 : }
3556 : :
3557 : : #define ST_SORT sort_checkpoint_bufferids
3558 : : #define ST_ELEMENT_TYPE CkptSortItem
3559 : : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3560 : : #define ST_SCOPE static
3561 : : #define ST_DEFINE
3562 : : #include "lib/sort_template.h"
3563 : :
3564 : : /*
3565 : : * BufferSync -- Write out all dirty buffers in the pool.
3566 : : *
3567 : : * This is called at checkpoint time to write out all dirty shared buffers.
3568 : : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3569 : : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3570 : : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3571 : : * even unlogged buffers, which are otherwise skipped. The remaining flags
3572 : : * currently have no effect here.
3573 : : */
3574 : : static void
3575 : 1950 : BufferSync(int flags)
3576 : : {
3577 : : uint64 buf_state;
3578 : : int buf_id;
3579 : : int num_to_scan;
3580 : : int num_spaces;
3581 : : int num_processed;
3582 : : int num_written;
3583 : 1950 : CkptTsStatus *per_ts_stat = NULL;
3584 : : Oid last_tsid;
3585 : : binaryheap *ts_heap;
3586 : : int i;
3587 : 1950 : uint64 mask = BM_DIRTY;
3588 : : WritebackContext wb_context;
3589 : :
3590 : : /*
3591 : : * Unless this is a shutdown checkpoint or we have been explicitly told,
3592 : : * we write only permanent, dirty buffers. But at shutdown or end of
3593 : : * recovery, we write all dirty buffers.
3594 : : */
3595 [ + + ]: 1950 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3596 : : CHECKPOINT_FLUSH_UNLOGGED))))
3597 : 1064 : mask |= BM_PERMANENT;
3598 : :
3599 : : /*
3600 : : * Loop over all buffers, and mark the ones that need to be written with
3601 : : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3602 : : * can estimate how much work needs to be done.
3603 : : *
3604 : : * This allows us to write only those pages that were dirty when the
3605 : : * checkpoint began, and not those that get dirtied while it proceeds.
3606 : : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3607 : : * later in this function, or by normal backends or the bgwriter cleaning
3608 : : * scan, the flag is cleared. Any buffer dirtied after this point won't
3609 : : * have the flag set.
3610 : : *
3611 : : * Note that if we fail to write some buffer, we may leave buffers with
3612 : : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3613 : : * certainly need to be written for the next checkpoint attempt, too.
3614 : : */
3615 : 1950 : num_to_scan = 0;
3616 [ + + ]: 13971310 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3617 : : {
3618 : 13969360 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3619 : 13969360 : uint64 set_bits = 0;
3620 : :
3621 : : /*
3622 : : * Header spinlock is enough to examine BM_DIRTY, see comment in
3623 : : * SyncOneBuffer.
3624 : : */
3625 : 13969360 : buf_state = LockBufHdr(bufHdr);
3626 : :
3627 [ + + ]: 13969360 : if ((buf_state & mask) == mask)
3628 : : {
3629 : : CkptSortItem *item;
3630 : :
3631 : 334783 : set_bits = BM_CHECKPOINT_NEEDED;
3632 : :
3633 : 334783 : item = &CkptBufferIds[num_to_scan++];
3634 : 334783 : item->buf_id = buf_id;
3635 : 334783 : item->tsId = bufHdr->tag.spcOid;
3636 : 334783 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3637 : 334783 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3638 : 334783 : item->blockNum = bufHdr->tag.blockNum;
3639 : : }
3640 : :
3641 : 13969360 : UnlockBufHdrExt(bufHdr, buf_state,
3642 : : set_bits, 0,
3643 : : 0);
3644 : :
3645 : : /* Check for barrier events in case NBuffers is large. */
3646 [ + + ]: 13969360 : if (ProcSignalBarrierPending)
3647 : 1 : ProcessProcSignalBarrier();
3648 : : }
3649 : :
3650 [ + + ]: 1950 : if (num_to_scan == 0)
3651 : 733 : return; /* nothing to do */
3652 : :
3653 : 1217 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3654 : :
3655 : : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3656 : :
3657 : : /*
3658 : : * Sort buffers that need to be written to reduce the likelihood of random
3659 : : * IO. The sorting is also important for the implementation of balancing
3660 : : * writes between tablespaces. Without balancing writes we'd potentially
3661 : : * end up writing to the tablespaces one-by-one; possibly overloading the
3662 : : * underlying system.
3663 : : */
3664 : 1217 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3665 : :
3666 : 1217 : num_spaces = 0;
3667 : :
3668 : : /*
3669 : : * Allocate progress status for each tablespace with buffers that need to
3670 : : * be flushed. This requires the to-be-flushed array to be sorted.
3671 : : */
3672 : 1217 : last_tsid = InvalidOid;
3673 [ + + ]: 336000 : for (i = 0; i < num_to_scan; i++)
3674 : : {
3675 : : CkptTsStatus *s;
3676 : : Oid cur_tsid;
3677 : :
3678 : 334783 : cur_tsid = CkptBufferIds[i].tsId;
3679 : :
3680 : : /*
3681 : : * Grow array of per-tablespace status structs, every time a new
3682 : : * tablespace is found.
3683 : : */
3684 [ + + + + ]: 334783 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3685 : 1840 : {
3686 : : Size sz;
3687 : :
3688 : 1840 : num_spaces++;
3689 : :
3690 : : /*
3691 : : * Not worth adding grow-by-power-of-2 logic here - even with a
3692 : : * few hundred tablespaces this should be fine.
3693 : : */
3694 : 1840 : sz = sizeof(CkptTsStatus) * num_spaces;
3695 : :
3696 [ + + ]: 1840 : if (per_ts_stat == NULL)
3697 : 1217 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3698 : : else
3699 : 623 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3700 : :
3701 : 1840 : s = &per_ts_stat[num_spaces - 1];
3702 : 1840 : memset(s, 0, sizeof(*s));
3703 : 1840 : s->tsId = cur_tsid;
3704 : :
3705 : : /*
3706 : : * The first buffer in this tablespace. As CkptBufferIds is sorted
3707 : : * by tablespace all (s->num_to_scan) buffers in this tablespace
3708 : : * will follow afterwards.
3709 : : */
3710 : 1840 : s->index = i;
3711 : :
3712 : : /*
3713 : : * progress_slice will be determined once we know how many buffers
3714 : : * are in each tablespace, i.e. after this loop.
3715 : : */
3716 : :
3717 : 1840 : last_tsid = cur_tsid;
3718 : : }
3719 : : else
3720 : : {
3721 : 332943 : s = &per_ts_stat[num_spaces - 1];
3722 : : }
3723 : :
3724 : 334783 : s->num_to_scan++;
3725 : :
3726 : : /* Check for barrier events. */
3727 [ - + ]: 334783 : if (ProcSignalBarrierPending)
3728 : 0 : ProcessProcSignalBarrier();
3729 : : }
3730 : :
3731 : : Assert(num_spaces > 0);
3732 : :
3733 : : /*
3734 : : * Build a min-heap over the write-progress in the individual tablespaces,
3735 : : * and compute how large a portion of the total progress a single
3736 : : * processed buffer is.
3737 : : */
3738 : 1217 : ts_heap = binaryheap_allocate(num_spaces,
3739 : : ts_ckpt_progress_comparator,
3740 : : NULL);
3741 : :
3742 [ + + ]: 3057 : for (i = 0; i < num_spaces; i++)
3743 : : {
3744 : 1840 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3745 : :
3746 : 1840 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3747 : :
3748 : 1840 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3749 : : }
3750 : :
3751 : 1217 : binaryheap_build(ts_heap);
3752 : :
3753 : : /*
3754 : : * Iterate through to-be-checkpointed buffers and write the ones (still)
3755 : : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3756 : : * tablespaces; otherwise the sorting would lead to only one tablespace
3757 : : * receiving writes at a time, making inefficient use of the hardware.
3758 : : */
3759 : 1217 : num_processed = 0;
3760 : 1217 : num_written = 0;
3761 [ + + ]: 336000 : while (!binaryheap_empty(ts_heap))
3762 : : {
3763 : 334783 : BufferDesc *bufHdr = NULL;
3764 : : CkptTsStatus *ts_stat = (CkptTsStatus *)
3765 : 334783 : DatumGetPointer(binaryheap_first(ts_heap));
3766 : :
3767 : 334783 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3768 : : Assert(buf_id != -1);
3769 : :
3770 : 334783 : bufHdr = GetBufferDescriptor(buf_id);
3771 : :
3772 : 334783 : num_processed++;
3773 : :
3774 : : /*
3775 : : * We don't need to acquire the lock here, because we're only looking
3776 : : * at a single bit. It's possible that someone else writes the buffer
3777 : : * and clears the flag right after we check, but that doesn't matter
3778 : : * since SyncOneBuffer will then do nothing. However, there is a
3779 : : * further race condition: it's conceivable that between the time we
3780 : : * examine the bit here and the time SyncOneBuffer acquires the lock,
3781 : : * someone else not only wrote the buffer but replaced it with another
3782 : : * page and dirtied it. In that improbable case, SyncOneBuffer will
3783 : : * write the buffer though we didn't need to. It doesn't seem worth
3784 : : * guarding against this, though.
3785 : : */
3786 [ + + ]: 334783 : if (pg_atomic_read_u64(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3787 : : {
3788 [ + - ]: 312654 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3789 : : {
3790 : : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3791 : 312654 : PendingCheckpointerStats.buffers_written++;
3792 : 312654 : num_written++;
3793 : : }
3794 : : }
3795 : :
3796 : : /*
3797 : : * Measure progress independent of actually having to flush the buffer
3798 : : * - otherwise writing become unbalanced.
3799 : : */
3800 : 334783 : ts_stat->progress += ts_stat->progress_slice;
3801 : 334783 : ts_stat->num_scanned++;
3802 : 334783 : ts_stat->index++;
3803 : :
3804 : : /* Have all the buffers from the tablespace been processed? */
3805 [ + + ]: 334783 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3806 : : {
3807 : 1840 : binaryheap_remove_first(ts_heap);
3808 : : }
3809 : : else
3810 : : {
3811 : : /* update heap with the new progress */
3812 : 332943 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3813 : : }
3814 : :
3815 : : /*
3816 : : * Sleep to throttle our I/O rate.
3817 : : *
3818 : : * (This will check for barrier events even if it doesn't sleep.)
3819 : : */
3820 : 334783 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3821 : : }
3822 : :
3823 : : /*
3824 : : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3825 : : * IOContext will always be IOCONTEXT_NORMAL.
3826 : : */
3827 : 1217 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3828 : :
3829 : 1217 : pfree(per_ts_stat);
3830 : 1217 : per_ts_stat = NULL;
3831 : 1217 : binaryheap_free(ts_heap);
3832 : :
3833 : : /*
3834 : : * Update checkpoint statistics. As noted above, this doesn't include
3835 : : * buffers written by other backends or bgwriter scan.
3836 : : */
3837 : 1217 : CheckpointStats.ckpt_bufs_written += num_written;
3838 : :
3839 : : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3840 : : }
3841 : :
3842 : : /*
3843 : : * BgBufferSync -- Write out some dirty buffers in the pool.
3844 : : *
3845 : : * This is called periodically by the background writer process.
3846 : : *
3847 : : * Returns true if it's appropriate for the bgwriter process to go into
3848 : : * low-power hibernation mode. (This happens if the strategy clock-sweep
3849 : : * has been "lapped" and no buffer allocations have occurred recently,
3850 : : * or if the bgwriter has been effectively disabled by setting
3851 : : * bgwriter_lru_maxpages to 0.)
3852 : : */
3853 : : bool
3854 : 15110 : BgBufferSync(WritebackContext *wb_context)
3855 : : {
3856 : : /* info obtained from freelist.c */
3857 : : int strategy_buf_id;
3858 : : uint32 strategy_passes;
3859 : : uint32 recent_alloc;
3860 : :
3861 : : /*
3862 : : * Information saved between calls so we can determine the strategy
3863 : : * point's advance rate and avoid scanning already-cleaned buffers.
3864 : : */
3865 : : static bool saved_info_valid = false;
3866 : : static int prev_strategy_buf_id;
3867 : : static uint32 prev_strategy_passes;
3868 : : static int next_to_clean;
3869 : : static uint32 next_passes;
3870 : :
3871 : : /* Moving averages of allocation rate and clean-buffer density */
3872 : : static float smoothed_alloc = 0;
3873 : : static float smoothed_density = 10.0;
3874 : :
3875 : : /* Potentially these could be tunables, but for now, not */
3876 : 15110 : float smoothing_samples = 16;
3877 : 15110 : float scan_whole_pool_milliseconds = 120000.0;
3878 : :
3879 : : /* Used to compute how far we scan ahead */
3880 : : long strategy_delta;
3881 : : int bufs_to_lap;
3882 : : int bufs_ahead;
3883 : : float scans_per_alloc;
3884 : : int reusable_buffers_est;
3885 : : int upcoming_alloc_est;
3886 : : int min_scan_buffers;
3887 : :
3888 : : /* Variables for the scanning loop proper */
3889 : : int num_to_scan;
3890 : : int num_written;
3891 : : int reusable_buffers;
3892 : :
3893 : : /* Variables for final smoothed_density update */
3894 : : long new_strategy_delta;
3895 : : uint32 new_recent_alloc;
3896 : :
3897 : : /*
3898 : : * Find out where the clock-sweep currently is, and how many buffer
3899 : : * allocations have happened since our last call.
3900 : : */
3901 : 15110 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3902 : :
3903 : : /* Report buffer alloc counts to pgstat */
3904 : 15110 : PendingBgWriterStats.buf_alloc += recent_alloc;
3905 : :
3906 : : /*
3907 : : * If we're not running the LRU scan, just stop after doing the stats
3908 : : * stuff. We mark the saved state invalid so that we can recover sanely
3909 : : * if LRU scan is turned back on later.
3910 : : */
3911 [ + + ]: 15110 : if (bgwriter_lru_maxpages <= 0)
3912 : : {
3913 : 39 : saved_info_valid = false;
3914 : 39 : return true;
3915 : : }
3916 : :
3917 : : /*
3918 : : * Compute strategy_delta = how many buffers have been scanned by the
3919 : : * clock-sweep since last time. If first time through, assume none. Then
3920 : : * see if we are still ahead of the clock-sweep, and if so, how many
3921 : : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3922 : : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3923 : : * behavior when the passes counts wrap around.
3924 : : */
3925 [ + + ]: 15071 : if (saved_info_valid)
3926 : : {
3927 : 14436 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3928 : :
3929 : 14436 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3930 : 14436 : strategy_delta += (long) passes_delta * NBuffers;
3931 : :
3932 : : Assert(strategy_delta >= 0);
3933 : :
3934 [ + + ]: 14436 : if ((int32) (next_passes - strategy_passes) > 0)
3935 : : {
3936 : : /* we're one pass ahead of the strategy point */
3937 : 2459 : bufs_to_lap = strategy_buf_id - next_to_clean;
3938 : : #ifdef BGW_DEBUG
3939 : : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3940 : : next_passes, next_to_clean,
3941 : : strategy_passes, strategy_buf_id,
3942 : : strategy_delta, bufs_to_lap);
3943 : : #endif
3944 : : }
3945 [ + + ]: 11977 : else if (next_passes == strategy_passes &&
3946 [ + + ]: 9380 : next_to_clean >= strategy_buf_id)
3947 : : {
3948 : : /* on same pass, but ahead or at least not behind */
3949 : 8415 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3950 : : #ifdef BGW_DEBUG
3951 : : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3952 : : next_passes, next_to_clean,
3953 : : strategy_passes, strategy_buf_id,
3954 : : strategy_delta, bufs_to_lap);
3955 : : #endif
3956 : : }
3957 : : else
3958 : : {
3959 : : /*
3960 : : * We're behind, so skip forward to the strategy point and start
3961 : : * cleaning from there.
3962 : : */
3963 : : #ifdef BGW_DEBUG
3964 : : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3965 : : next_passes, next_to_clean,
3966 : : strategy_passes, strategy_buf_id,
3967 : : strategy_delta);
3968 : : #endif
3969 : 3562 : next_to_clean = strategy_buf_id;
3970 : 3562 : next_passes = strategy_passes;
3971 : 3562 : bufs_to_lap = NBuffers;
3972 : : }
3973 : : }
3974 : : else
3975 : : {
3976 : : /*
3977 : : * Initializing at startup or after LRU scanning had been off. Always
3978 : : * start at the strategy point.
3979 : : */
3980 : : #ifdef BGW_DEBUG
3981 : : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3982 : : strategy_passes, strategy_buf_id);
3983 : : #endif
3984 : 635 : strategy_delta = 0;
3985 : 635 : next_to_clean = strategy_buf_id;
3986 : 635 : next_passes = strategy_passes;
3987 : 635 : bufs_to_lap = NBuffers;
3988 : : }
3989 : :
3990 : : /* Update saved info for next time */
3991 : 15071 : prev_strategy_buf_id = strategy_buf_id;
3992 : 15071 : prev_strategy_passes = strategy_passes;
3993 : 15071 : saved_info_valid = true;
3994 : :
3995 : : /*
3996 : : * Compute how many buffers had to be scanned for each new allocation, ie,
3997 : : * 1/density of reusable buffers, and track a moving average of that.
3998 : : *
3999 : : * If the strategy point didn't move, we don't update the density estimate
4000 : : */
4001 [ + + + - ]: 15071 : if (strategy_delta > 0 && recent_alloc > 0)
4002 : : {
4003 : 8175 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
4004 : 8175 : smoothed_density += (scans_per_alloc - smoothed_density) /
4005 : : smoothing_samples;
4006 : : }
4007 : :
4008 : : /*
4009 : : * Estimate how many reusable buffers there are between the current
4010 : : * strategy point and where we've scanned ahead to, based on the smoothed
4011 : : * density estimate.
4012 : : */
4013 : 15071 : bufs_ahead = NBuffers - bufs_to_lap;
4014 : 15071 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
4015 : :
4016 : : /*
4017 : : * Track a moving average of recent buffer allocations. Here, rather than
4018 : : * a true average we want a fast-attack, slow-decline behavior: we
4019 : : * immediately follow any increase.
4020 : : */
4021 [ + + ]: 15071 : if (smoothed_alloc <= (float) recent_alloc)
4022 : 3939 : smoothed_alloc = recent_alloc;
4023 : : else
4024 : 11132 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
4025 : : smoothing_samples;
4026 : :
4027 : : /* Scale the estimate by a GUC to allow more aggressive tuning. */
4028 : 15071 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
4029 : :
4030 : : /*
4031 : : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
4032 : : * eventually underflow to zero, and the underflows produce annoying
4033 : : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
4034 : : * zero, there's no point in tracking smaller and smaller values of
4035 : : * smoothed_alloc, so just reset it to exactly zero to avoid this
4036 : : * syndrome. It will pop back up as soon as recent_alloc increases.
4037 : : */
4038 [ + + ]: 15071 : if (upcoming_alloc_est == 0)
4039 : 2282 : smoothed_alloc = 0;
4040 : :
4041 : : /*
4042 : : * Even in cases where there's been little or no buffer allocation
4043 : : * activity, we want to make a small amount of progress through the buffer
4044 : : * cache so that as many reusable buffers as possible are clean after an
4045 : : * idle period.
4046 : : *
4047 : : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
4048 : : * the BGW will be called during the scan_whole_pool time; slice the
4049 : : * buffer pool into that many sections.
4050 : : */
4051 : 15071 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
4052 : :
4053 [ + + ]: 15071 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
4054 : : {
4055 : : #ifdef BGW_DEBUG
4056 : : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
4057 : : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
4058 : : #endif
4059 : 7184 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
4060 : : }
4061 : :
4062 : : /*
4063 : : * Now write out dirty reusable buffers, working forward from the
4064 : : * next_to_clean point, until we have lapped the strategy scan, or cleaned
4065 : : * enough buffers to match our estimate of the next cycle's allocation
4066 : : * requirements, or hit the bgwriter_lru_maxpages limit.
4067 : : */
4068 : :
4069 : 15071 : num_to_scan = bufs_to_lap;
4070 : 15071 : num_written = 0;
4071 : 15071 : reusable_buffers = reusable_buffers_est;
4072 : :
4073 : : /* Execute the LRU scan */
4074 [ + + + + ]: 2058088 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
4075 : : {
4076 : 2043020 : int sync_state = SyncOneBuffer(next_to_clean, true,
4077 : : wb_context);
4078 : :
4079 [ + + ]: 2043020 : if (++next_to_clean >= NBuffers)
4080 : : {
4081 : 3417 : next_to_clean = 0;
4082 : 3417 : next_passes++;
4083 : : }
4084 : 2043020 : num_to_scan--;
4085 : :
4086 [ + + ]: 2043020 : if (sync_state & BUF_WRITTEN)
4087 : : {
4088 : 27982 : reusable_buffers++;
4089 [ + + ]: 27982 : if (++num_written >= bgwriter_lru_maxpages)
4090 : : {
4091 : 3 : PendingBgWriterStats.maxwritten_clean++;
4092 : 3 : break;
4093 : : }
4094 : : }
4095 [ + + ]: 2015038 : else if (sync_state & BUF_REUSABLE)
4096 : 1579077 : reusable_buffers++;
4097 : : }
4098 : :
4099 : 15071 : PendingBgWriterStats.buf_written_clean += num_written;
4100 : :
4101 : : #ifdef BGW_DEBUG
4102 : : elog(DEBUG1, "bgwriter: recent_alloc=%u smoothed=%.2f delta=%ld ahead=%d density=%.2f reusable_est=%d upcoming_est=%d scanned=%d wrote=%d reusable=%d",
4103 : : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
4104 : : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
4105 : : bufs_to_lap - num_to_scan,
4106 : : num_written,
4107 : : reusable_buffers - reusable_buffers_est);
4108 : : #endif
4109 : :
4110 : : /*
4111 : : * Consider the above scan as being like a new allocation scan.
4112 : : * Characterize its density and update the smoothed one based on it. This
4113 : : * effectively halves the moving average period in cases where both the
4114 : : * strategy and the background writer are doing some useful scanning,
4115 : : * which is helpful because a long memory isn't as desirable on the
4116 : : * density estimates.
4117 : : */
4118 : 15071 : new_strategy_delta = bufs_to_lap - num_to_scan;
4119 : 15071 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
4120 [ + + + + ]: 15071 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
4121 : : {
4122 : 12872 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
4123 : 12872 : smoothed_density += (scans_per_alloc - smoothed_density) /
4124 : : smoothing_samples;
4125 : :
4126 : : #ifdef BGW_DEBUG
4127 : : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
4128 : : new_recent_alloc, new_strategy_delta,
4129 : : scans_per_alloc, smoothed_density);
4130 : : #endif
4131 : : }
4132 : :
4133 : : /* Return true if OK to hibernate */
4134 [ + + + - ]: 15071 : return (bufs_to_lap == 0 && recent_alloc == 0);
4135 : : }
4136 : :
4137 : : /*
4138 : : * SyncOneBuffer -- process a single buffer during syncing.
4139 : : *
4140 : : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
4141 : : * buffers marked recently used, as these are not replacement candidates.
4142 : : *
4143 : : * Returns a bitmask containing the following flag bits:
4144 : : * BUF_WRITTEN: we wrote the buffer.
4145 : : * BUF_REUSABLE: buffer is available for replacement, ie, it has
4146 : : * pin count 0 and usage count 0.
4147 : : *
4148 : : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
4149 : : * after locking it, but we don't care all that much.)
4150 : : */
4151 : : static int
4152 : 2355674 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
4153 : : {
4154 : 2355674 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
4155 : 2355674 : int result = 0;
4156 : : uint64 buf_state;
4157 : : BufferTag tag;
4158 : :
4159 : : /* Make sure we can handle the pin */
4160 : 2355674 : ReservePrivateRefCountEntry();
4161 : 2355674 : ResourceOwnerEnlarge(CurrentResourceOwner);
4162 : :
4163 : : /*
4164 : : * Check whether buffer needs writing.
4165 : : *
4166 : : * We can make this check without taking the buffer content lock so long
4167 : : * as we mark pages dirty in access methods *before* logging changes with
4168 : : * XLogInsert(): if someone marks the buffer dirty just after our check we
4169 : : * don't worry because our checkpoint.redo points before log record for
4170 : : * upcoming changes and so we are not required to write such dirty buffer.
4171 : : */
4172 : 2355674 : buf_state = LockBufHdr(bufHdr);
4173 : :
4174 [ + + ]: 2355674 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
4175 [ + + ]: 2352355 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
4176 : : {
4177 : 1609254 : result |= BUF_REUSABLE;
4178 : : }
4179 [ + + ]: 746420 : else if (skip_recently_used)
4180 : : {
4181 : : /* Caller told us not to write recently-used buffers */
4182 : 435961 : UnlockBufHdr(bufHdr);
4183 : 435961 : return result;
4184 : : }
4185 : :
4186 [ + + + + ]: 1919713 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
4187 : : {
4188 : : /* It's clean, so nothing to do */
4189 : 1579077 : UnlockBufHdr(bufHdr);
4190 : 1579077 : return result;
4191 : : }
4192 : :
4193 : : /*
4194 : : * Pin it, share-exclusive-lock it, write it. (FlushBuffer will do
4195 : : * nothing if the buffer is clean by the time we've locked it.)
4196 : : */
4197 : 340636 : PinBuffer_Locked(bufHdr);
4198 : :
4199 : 340636 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4200 : :
4201 : 340636 : tag = bufHdr->tag;
4202 : :
4203 : 340636 : UnpinBuffer(bufHdr);
4204 : :
4205 : : /*
4206 : : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
4207 : : * IOContext will always be IOCONTEXT_NORMAL.
4208 : : */
4209 : 340636 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
4210 : :
4211 : 340636 : return result | BUF_WRITTEN;
4212 : : }
4213 : :
4214 : : /*
4215 : : * AtEOXact_Buffers - clean up at end of transaction.
4216 : : *
4217 : : * As of PostgreSQL 8.0, buffer pins should get released by the
4218 : : * ResourceOwner mechanism. This routine is just a debugging
4219 : : * cross-check that no pins remain.
4220 : : */
4221 : : void
4222 : 654586 : AtEOXact_Buffers(bool isCommit)
4223 : : {
4224 : 654586 : CheckForBufferLeaks();
4225 : :
4226 : 654586 : AtEOXact_LocalBuffers(isCommit);
4227 : :
4228 : : Assert(PrivateRefCountOverflowed == 0);
4229 : 654586 : }
4230 : :
4231 : : /*
4232 : : * Initialize access to shared buffer pool
4233 : : *
4234 : : * This is called during backend startup (whether standalone or under the
4235 : : * postmaster). It sets up for this backend's access to the already-existing
4236 : : * buffer pool.
4237 : : */
4238 : : void
4239 : 24803 : InitBufferManagerAccess(void)
4240 : : {
4241 : : /*
4242 : : * An advisory limit on the number of pins each backend should hold, based
4243 : : * on shared_buffers and the maximum number of connections possible.
4244 : : * That's very pessimistic, but outside toy-sized shared_buffers it should
4245 : : * allow plenty of pins. LimitAdditionalPins() and
4246 : : * GetAdditionalPinLimit() can be used to check the remaining balance.
4247 : : */
4248 : 24803 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4249 : :
4250 : 24803 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4251 : 24803 : memset(&PrivateRefCountArrayKeys, 0, sizeof(PrivateRefCountArrayKeys));
4252 : :
4253 : 24803 : PrivateRefCountHash = refcount_create(CurrentMemoryContext, 100, NULL);
4254 : :
4255 : : /*
4256 : : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4257 : : * the corresponding phase of backend shutdown.
4258 : : */
4259 : : Assert(MyProc != NULL);
4260 : 24803 : on_shmem_exit(AtProcExit_Buffers, 0);
4261 : 24803 : }
4262 : :
4263 : : /*
4264 : : * During backend exit, ensure that we released all shared-buffer locks and
4265 : : * assert that we have no remaining pins.
4266 : : */
4267 : : static void
4268 : 24803 : AtProcExit_Buffers(int code, Datum arg)
4269 : : {
4270 : 24803 : UnlockBuffers();
4271 : :
4272 : 24803 : CheckForBufferLeaks();
4273 : :
4274 : : /* localbuf.c needs a chance too */
4275 : 24803 : AtProcExit_LocalBuffers();
4276 : 24803 : }
4277 : :
4278 : : /*
4279 : : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4280 : : *
4281 : : * As of PostgreSQL 8.0, buffer pins should get released by the
4282 : : * ResourceOwner mechanism. This routine is just a debugging
4283 : : * cross-check that no pins remain.
4284 : : */
4285 : : static void
4286 : 679389 : CheckForBufferLeaks(void)
4287 : : {
4288 : : #ifdef USE_ASSERT_CHECKING
4289 : : int RefCountErrors = 0;
4290 : : PrivateRefCountEntry *res;
4291 : : int i;
4292 : : char *s;
4293 : :
4294 : : /* check the array */
4295 : : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4296 : : {
4297 : : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4298 : : {
4299 : : res = &PrivateRefCountArray[i];
4300 : :
4301 : : s = DebugPrintBufferRefcount(res->buffer);
4302 : : elog(WARNING, "buffer refcount leak: %s", s);
4303 : : pfree(s);
4304 : :
4305 : : RefCountErrors++;
4306 : : }
4307 : : }
4308 : :
4309 : : /* if necessary search the hash */
4310 : : if (PrivateRefCountOverflowed)
4311 : : {
4312 : : refcount_iterator iter;
4313 : :
4314 : : refcount_start_iterate(PrivateRefCountHash, &iter);
4315 : : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4316 : : {
4317 : : s = DebugPrintBufferRefcount(res->buffer);
4318 : : elog(WARNING, "buffer refcount leak: %s", s);
4319 : : pfree(s);
4320 : : RefCountErrors++;
4321 : : }
4322 : : }
4323 : :
4324 : : Assert(RefCountErrors == 0);
4325 : : #endif
4326 : 679389 : }
4327 : :
4328 : : #ifdef USE_ASSERT_CHECKING
4329 : : /*
4330 : : * Check for exclusive-locked catalog buffers. This is the core of
4331 : : * AssertCouldGetRelation().
4332 : : *
4333 : : * A backend would self-deadlock on the content lock if the catalog scan read
4334 : : * the exclusive-locked buffer. The main threat is exclusive-locked buffers
4335 : : * of catalogs used in relcache, because a catcache search on any catalog may
4336 : : * build that catalog's relcache entry. We don't have an inventory of
4337 : : * catalogs relcache uses, so just check buffers of most catalogs.
4338 : : *
4339 : : * It's better to minimize waits while holding an exclusive buffer lock, so it
4340 : : * would be nice to broaden this check not to be catalog-specific. However,
4341 : : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4342 : : * read tables. That is deadlock-free as long as there's no loop in the
4343 : : * dependency graph: modifying table A may cause an opclass to read table B,
4344 : : * but it must not cause a read of table A.
4345 : : */
4346 : : void
4347 : : AssertBufferLocksPermitCatalogRead(void)
4348 : : {
4349 : : PrivateRefCountEntry *res;
4350 : :
4351 : : /* check the array */
4352 : : for (int i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4353 : : {
4354 : : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4355 : : {
4356 : : res = &PrivateRefCountArray[i];
4357 : :
4358 : : if (res->buffer == InvalidBuffer)
4359 : : continue;
4360 : :
4361 : : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4362 : : }
4363 : : }
4364 : :
4365 : : /* if necessary search the hash */
4366 : : if (PrivateRefCountOverflowed)
4367 : : {
4368 : : refcount_iterator iter;
4369 : :
4370 : : refcount_start_iterate(PrivateRefCountHash, &iter);
4371 : : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4372 : : {
4373 : : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4374 : : }
4375 : : }
4376 : : }
4377 : :
4378 : : static void
4379 : : AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode)
4380 : : {
4381 : : BufferDesc *bufHdr = GetBufferDescriptor(buffer - 1);
4382 : : BufferTag tag;
4383 : : Oid relid;
4384 : :
4385 : : if (mode != BUFFER_LOCK_EXCLUSIVE)
4386 : : return;
4387 : :
4388 : : tag = bufHdr->tag;
4389 : :
4390 : : /*
4391 : : * This relNumber==relid assumption holds until a catalog experiences
4392 : : * VACUUM FULL or similar. After a command like that, relNumber will be
4393 : : * in the normal (non-catalog) range, and we lose the ability to detect
4394 : : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4395 : : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4396 : : * held lock.
4397 : : */
4398 : : relid = tag.relNumber;
4399 : :
4400 : : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4401 : : return;
4402 : :
4403 : : Assert(!IsCatalogRelationOid(relid));
4404 : : }
4405 : : #endif
4406 : :
4407 : :
4408 : : /*
4409 : : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4410 : : */
4411 : : char *
4412 : 46 : DebugPrintBufferRefcount(Buffer buffer)
4413 : : {
4414 : : BufferDesc *buf;
4415 : : int32 loccount;
4416 : : char *result;
4417 : : ProcNumber backend;
4418 : : uint64 buf_state;
4419 : :
4420 : : Assert(BufferIsValid(buffer));
4421 [ + + ]: 46 : if (BufferIsLocal(buffer))
4422 : : {
4423 : 16 : buf = GetLocalBufferDescriptor(-buffer - 1);
4424 : 16 : loccount = LocalRefCount[-buffer - 1];
4425 : 16 : backend = MyProcNumber;
4426 : : }
4427 : : else
4428 : : {
4429 : 30 : buf = GetBufferDescriptor(buffer - 1);
4430 : 30 : loccount = GetPrivateRefCount(buffer);
4431 : 30 : backend = INVALID_PROC_NUMBER;
4432 : : }
4433 : :
4434 : : /* theoretically we should lock the bufHdr here */
4435 : 46 : buf_state = pg_atomic_read_u64(&buf->state);
4436 : :
4437 : 46 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%" PRIx64 ", refcount=%u %d)",
4438 : : buffer,
4439 : 46 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4440 : : BufTagGetForkNum(&buf->tag)).str,
4441 : : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4442 : : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4443 : 46 : return result;
4444 : : }
4445 : :
4446 : : /*
4447 : : * CheckPointBuffers
4448 : : *
4449 : : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4450 : : *
4451 : : * Note: temporary relations do not participate in checkpoints, so they don't
4452 : : * need to be flushed.
4453 : : */
4454 : : void
4455 : 1950 : CheckPointBuffers(int flags)
4456 : : {
4457 : 1950 : BufferSync(flags);
4458 : 1950 : }
4459 : :
4460 : : /*
4461 : : * BufferGetBlockNumber
4462 : : * Returns the block number associated with a buffer.
4463 : : *
4464 : : * Note:
4465 : : * Assumes that the buffer is valid and pinned, else the
4466 : : * value may be obsolete immediately...
4467 : : */
4468 : : BlockNumber
4469 : 80744448 : BufferGetBlockNumber(Buffer buffer)
4470 : : {
4471 : : BufferDesc *bufHdr;
4472 : :
4473 : : Assert(BufferIsPinned(buffer));
4474 : :
4475 [ + + ]: 80744448 : if (BufferIsLocal(buffer))
4476 : 2561323 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4477 : : else
4478 : 78183125 : bufHdr = GetBufferDescriptor(buffer - 1);
4479 : :
4480 : : /* pinned, so OK to read tag without spinlock */
4481 : 80744448 : return bufHdr->tag.blockNum;
4482 : : }
4483 : :
4484 : : /*
4485 : : * BufferGetTag
4486 : : * Returns the relfilelocator, fork number and block number associated with
4487 : : * a buffer.
4488 : : */
4489 : : void
4490 : 27662970 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4491 : : BlockNumber *blknum)
4492 : : {
4493 : : BufferDesc *bufHdr;
4494 : :
4495 : : /* Do the same checks as BufferGetBlockNumber. */
4496 : : Assert(BufferIsPinned(buffer));
4497 : :
4498 [ - + ]: 27662970 : if (BufferIsLocal(buffer))
4499 : 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4500 : : else
4501 : 27662970 : bufHdr = GetBufferDescriptor(buffer - 1);
4502 : :
4503 : : /* pinned, so OK to read tag without spinlock */
4504 : 27662970 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4505 : 27662970 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4506 : 27662970 : *blknum = bufHdr->tag.blockNum;
4507 : 27662970 : }
4508 : :
4509 : : /*
4510 : : * FlushBuffer
4511 : : * Physically write out a shared buffer.
4512 : : *
4513 : : * NOTE: this actually just passes the buffer contents to the kernel; the
4514 : : * real write to disk won't happen until the kernel feels like it. This
4515 : : * is okay from our point of view since we can redo the changes from WAL.
4516 : : * However, we will need to force the changes to disk via fsync before
4517 : : * we can checkpoint WAL.
4518 : : *
4519 : : * The caller must hold a pin on the buffer and have
4520 : : * (share-)exclusively-locked the buffer contents.
4521 : : *
4522 : : * If the caller has an smgr reference for the buffer's relation, pass it
4523 : : * as the second parameter. If not, pass NULL.
4524 : : */
4525 : : static void
4526 : 700411 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4527 : : IOContext io_context)
4528 : : {
4529 : : XLogRecPtr recptr;
4530 : : ErrorContextCallback errcallback;
4531 : : instr_time io_start;
4532 : : Block bufBlock;
4533 : :
4534 : : Assert(BufferLockHeldByMeInMode(buf, BUFFER_LOCK_EXCLUSIVE) ||
4535 : : BufferLockHeldByMeInMode(buf, BUFFER_LOCK_SHARE_EXCLUSIVE));
4536 : :
4537 : : /*
4538 : : * Try to start an I/O operation. If StartBufferIO returns false, then
4539 : : * someone else flushed the buffer before we could, so we need not do
4540 : : * anything.
4541 : : */
4542 [ + + ]: 700411 : if (StartSharedBufferIO(buf, false, true, NULL) == BUFFER_IO_ALREADY_DONE)
4543 : 10 : return;
4544 : :
4545 : : /* Setup error traceback support for ereport() */
4546 : 700401 : errcallback.callback = shared_buffer_write_error_callback;
4547 : 700401 : errcallback.arg = buf;
4548 : 700401 : errcallback.previous = error_context_stack;
4549 : 700401 : error_context_stack = &errcallback;
4550 : :
4551 : : /* Find smgr relation for buffer */
4552 [ + + ]: 700401 : if (reln == NULL)
4553 : 697802 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4554 : :
4555 : : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4556 : : buf->tag.blockNum,
4557 : : reln->smgr_rlocator.locator.spcOid,
4558 : : reln->smgr_rlocator.locator.dbOid,
4559 : : reln->smgr_rlocator.locator.relNumber);
4560 : :
4561 : : /*
4562 : : * As we hold at least a share-exclusive lock on the buffer, the LSN
4563 : : * cannot change during the flush (and thus can't be torn).
4564 : : */
4565 : 700401 : recptr = BufferGetLSN(buf);
4566 : :
4567 : : /*
4568 : : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4569 : : * rule that log updates must hit disk before any of the data-file changes
4570 : : * they describe do.
4571 : : *
4572 : : * However, this rule does not apply to unlogged relations, which will be
4573 : : * lost after a crash anyway. Most unlogged relation pages do not bear
4574 : : * LSNs since we never emit WAL records for them, and therefore flushing
4575 : : * up through the buffer LSN would be useless, but harmless. However,
4576 : : * some index AMs use LSNs internally to detect concurrent page
4577 : : * modifications, and therefore unlogged index pages bear "fake" LSNs
4578 : : * generated by XLogGetFakeLSN. It is unlikely but possible that the fake
4579 : : * LSN counter could advance past the WAL insertion point; and if it did
4580 : : * happen, attempting to flush WAL through that location would fail, with
4581 : : * disastrous system-wide consequences. To make sure that can't happen,
4582 : : * skip the flush if the buffer isn't permanent.
4583 : : */
4584 [ + + ]: 700401 : if (pg_atomic_read_u64(&buf->state) & BM_PERMANENT)
4585 : 698513 : XLogFlush(recptr);
4586 : :
4587 : : /*
4588 : : * Now it's safe to write the buffer to disk. Note that no one else should
4589 : : * have been able to write it, while we were busy with log flushing,
4590 : : * because we got the exclusive right to perform I/O by setting the
4591 : : * BM_IO_IN_PROGRESS bit.
4592 : : */
4593 : 700401 : bufBlock = BufHdrGetBlock(buf);
4594 : :
4595 : : /* Update page checksum if desired. */
4596 : 700401 : PageSetChecksum((Page) bufBlock, buf->tag.blockNum);
4597 : :
4598 : 700401 : io_start = pgstat_prepare_io_time(track_io_timing);
4599 : :
4600 : 700401 : smgrwrite(reln,
4601 : 700401 : BufTagGetForkNum(&buf->tag),
4602 : : buf->tag.blockNum,
4603 : : bufBlock,
4604 : : false);
4605 : :
4606 : : /*
4607 : : * When a strategy is in use, only flushes of dirty buffers already in the
4608 : : * strategy ring are counted as strategy writes (IOCONTEXT
4609 : : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4610 : : * statistics tracking.
4611 : : *
4612 : : * If a shared buffer initially added to the ring must be flushed before
4613 : : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4614 : : *
4615 : : * If a shared buffer which was added to the ring later because the
4616 : : * current strategy buffer is pinned or in use or because all strategy
4617 : : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4618 : : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4619 : : * (from_ring will be false).
4620 : : *
4621 : : * When a strategy is not in use, the write can only be a "regular" write
4622 : : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4623 : : */
4624 : 700401 : pgstat_count_io_op_time(io_object, io_context,
4625 : : IOOP_WRITE, io_start, 1, BLCKSZ);
4626 : :
4627 : 700401 : pgBufferUsage.shared_blks_written++;
4628 : :
4629 : : /*
4630 : : * Mark the buffer as clean and end the BM_IO_IN_PROGRESS state.
4631 : : */
4632 : 700401 : TerminateBufferIO(buf, true, 0, true, false);
4633 : :
4634 : : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4635 : : buf->tag.blockNum,
4636 : : reln->smgr_rlocator.locator.spcOid,
4637 : : reln->smgr_rlocator.locator.dbOid,
4638 : : reln->smgr_rlocator.locator.relNumber);
4639 : :
4640 : : /* Pop the error context stack */
4641 : 700401 : error_context_stack = errcallback.previous;
4642 : : }
4643 : :
4644 : : /*
4645 : : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4646 : : * before/after calling FlushBuffer().
4647 : : */
4648 : : static void
4649 : 344339 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4650 : : IOObject io_object, IOContext io_context)
4651 : : {
4652 : 344339 : Buffer buffer = BufferDescriptorGetBuffer(buf);
4653 : :
4654 : 344339 : BufferLockAcquire(buffer, buf, BUFFER_LOCK_SHARE_EXCLUSIVE);
4655 : 344339 : FlushBuffer(buf, reln, io_object, io_context);
4656 : 344339 : BufferLockUnlock(buffer, buf);
4657 : 344339 : }
4658 : :
4659 : : /*
4660 : : * RelationGetNumberOfBlocksInFork
4661 : : * Determines the current number of pages in the specified relation fork.
4662 : : *
4663 : : * Note that the accuracy of the result will depend on the details of the
4664 : : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4665 : : * it might not be.
4666 : : */
4667 : : BlockNumber
4668 : 2381931 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4669 : : {
4670 [ + + + + : 2381931 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
+ + ]
4671 : : {
4672 : : /*
4673 : : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4674 : : * tableam returns the size in bytes - but for the purpose of this
4675 : : * routine, we want the number of blocks. Therefore divide, rounding
4676 : : * up.
4677 : : */
4678 : : uint64 szbytes;
4679 : :
4680 : 1703814 : szbytes = table_relation_size(relation, forkNum);
4681 : :
4682 : 1703795 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4683 : : }
4684 [ + - + + : 678117 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
- + - - -
- ]
4685 : : {
4686 : 678117 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4687 : : }
4688 : : else
4689 : : Assert(false);
4690 : :
4691 : 0 : return 0; /* keep compiler quiet */
4692 : : }
4693 : :
4694 : : /*
4695 : : * BufferIsPermanent
4696 : : * Determines whether a buffer will potentially still be around after
4697 : : * a crash. Caller must hold a buffer pin.
4698 : : */
4699 : : bool
4700 : 17305037 : BufferIsPermanent(Buffer buffer)
4701 : : {
4702 : : BufferDesc *bufHdr;
4703 : :
4704 : : /* Local buffers are used only for temp relations. */
4705 [ + + ]: 17305037 : if (BufferIsLocal(buffer))
4706 : 828431 : return false;
4707 : :
4708 : : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4709 : : Assert(BufferIsValid(buffer));
4710 : : Assert(BufferIsPinned(buffer));
4711 : :
4712 : : /*
4713 : : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4714 : : * need not bother with the buffer header spinlock. Even if someone else
4715 : : * changes the buffer header state while we're doing this, the state is
4716 : : * changed atomically, so we'll read the old value or the new value, but
4717 : : * not random garbage.
4718 : : */
4719 : 16476606 : bufHdr = GetBufferDescriptor(buffer - 1);
4720 : 16476606 : return (pg_atomic_read_u64(&bufHdr->state) & BM_PERMANENT) != 0;
4721 : : }
4722 : :
4723 : : /*
4724 : : * BufferGetLSNAtomic
4725 : : * Retrieves the LSN of the buffer atomically.
4726 : : *
4727 : : * This is necessary for some callers who may only hold a share lock on
4728 : : * the buffer. A share lock allows a concurrent backend to set hint bits
4729 : : * on the page, which in turn may require a WAL record to be emitted.
4730 : : *
4731 : : * On platforms with 8 byte atomic reads/writes, we don't need to do any
4732 : : * additional locking. On platforms not supporting such 8 byte atomic
4733 : : * reads/writes, we need to actually take the header lock.
4734 : : */
4735 : : XLogRecPtr
4736 : 8930619 : BufferGetLSNAtomic(Buffer buffer)
4737 : : {
4738 : : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4739 : : Assert(BufferIsValid(buffer));
4740 : : Assert(BufferIsPinned(buffer));
4741 : :
4742 : : #ifdef PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
4743 : 8930619 : return PageGetLSN(BufferGetPage(buffer));
4744 : : #else
4745 : : {
4746 : : char *page = BufferGetPage(buffer);
4747 : : BufferDesc *bufHdr;
4748 : : XLogRecPtr lsn;
4749 : :
4750 : : /*
4751 : : * If we don't need locking for correctness, fastpath out.
4752 : : */
4753 : : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4754 : : return PageGetLSN(page);
4755 : :
4756 : : bufHdr = GetBufferDescriptor(buffer - 1);
4757 : : LockBufHdr(bufHdr);
4758 : : lsn = PageGetLSN(page);
4759 : : UnlockBufHdr(bufHdr);
4760 : :
4761 : : return lsn;
4762 : : }
4763 : : #endif
4764 : : }
4765 : :
4766 : : /* ---------------------------------------------------------------------
4767 : : * DropRelationBuffers
4768 : : *
4769 : : * This function removes from the buffer pool all the pages of the
4770 : : * specified relation forks that have block numbers >= firstDelBlock.
4771 : : * (In particular, with firstDelBlock = 0, all pages are removed.)
4772 : : * Dirty pages are simply dropped, without bothering to write them
4773 : : * out first. Therefore, this is NOT rollback-able, and so should be
4774 : : * used only with extreme caution!
4775 : : *
4776 : : * Currently, this is called only from smgr.c when the underlying file
4777 : : * is about to be deleted or truncated (firstDelBlock is needed for
4778 : : * the truncation case). The data in the affected pages would therefore
4779 : : * be deleted momentarily anyway, and there is no point in writing it.
4780 : : * It is the responsibility of higher-level code to ensure that the
4781 : : * deletion or truncation does not lose any data that could be needed
4782 : : * later. It is also the responsibility of higher-level code to ensure
4783 : : * that no other process could be trying to load more pages of the
4784 : : * relation into buffers.
4785 : : * --------------------------------------------------------------------
4786 : : */
4787 : : void
4788 : 788 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4789 : : int nforks, BlockNumber *firstDelBlock)
4790 : : {
4791 : : int i;
4792 : : int j;
4793 : : RelFileLocatorBackend rlocator;
4794 : : BlockNumber nForkBlock[MAX_FORKNUM];
4795 : 788 : uint64 nBlocksToInvalidate = 0;
4796 : :
4797 : 788 : rlocator = smgr_reln->smgr_rlocator;
4798 : :
4799 : : /* If it's a local relation, it's localbuf.c's problem. */
4800 [ + + ]: 788 : if (RelFileLocatorBackendIsTemp(rlocator))
4801 : : {
4802 [ + - ]: 498 : if (rlocator.backend == MyProcNumber)
4803 : 498 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4804 : : firstDelBlock);
4805 : :
4806 : 537 : return;
4807 : : }
4808 : :
4809 : : /*
4810 : : * To remove all the pages of the specified relation forks from the buffer
4811 : : * pool, we need to scan the entire buffer pool but we can optimize it by
4812 : : * finding the buffers from BufMapping table provided we know the exact
4813 : : * size of each fork of the relation. The exact size is required to ensure
4814 : : * that we don't leave any buffer for the relation being dropped as
4815 : : * otherwise the background writer or checkpointer can lead to a PANIC
4816 : : * error while flushing buffers corresponding to files that don't exist.
4817 : : *
4818 : : * To know the exact size, we rely on the size cached for each fork by us
4819 : : * during recovery which limits the optimization to recovery and on
4820 : : * standbys but we can easily extend it once we have shared cache for
4821 : : * relation size.
4822 : : *
4823 : : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4824 : : * and the future writes keeps the cached value up-to-date. See
4825 : : * smgrextend. It is possible that the value of the first lseek is smaller
4826 : : * than the actual number of existing blocks in the file due to buggy
4827 : : * Linux kernels that might not have accounted for the recent write. But
4828 : : * that should be fine because there must not be any buffers after that
4829 : : * file size.
4830 : : */
4831 [ + + ]: 390 : for (i = 0; i < nforks; i++)
4832 : : {
4833 : : /* Get the number of blocks for a relation's fork */
4834 : 338 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4835 : :
4836 [ + + ]: 338 : if (nForkBlock[i] == InvalidBlockNumber)
4837 : : {
4838 : 238 : nBlocksToInvalidate = InvalidBlockNumber;
4839 : 238 : break;
4840 : : }
4841 : :
4842 : : /* calculate the number of blocks to be invalidated */
4843 : 100 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4844 : : }
4845 : :
4846 : : /*
4847 : : * We apply the optimization iff the total number of blocks to invalidate
4848 : : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4849 : : */
4850 [ + + ]: 290 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4851 [ + + ]: 52 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4852 : : {
4853 [ + + ]: 107 : for (j = 0; j < nforks; j++)
4854 : 68 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4855 : 68 : nForkBlock[j], firstDelBlock[j]);
4856 : 39 : return;
4857 : : }
4858 : :
4859 [ + + ]: 3348603 : for (i = 0; i < NBuffers; i++)
4860 : : {
4861 : 3348352 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4862 : :
4863 : : /*
4864 : : * We can make this a tad faster by prechecking the buffer tag before
4865 : : * we attempt to lock the buffer; this saves a lot of lock
4866 : : * acquisitions in typical cases. It should be safe because the
4867 : : * caller must have AccessExclusiveLock on the relation, or some other
4868 : : * reason to be certain that no one is loading new pages of the rel
4869 : : * into the buffer pool. (Otherwise we might well miss such pages
4870 : : * entirely.) Therefore, while the tag might be changing while we
4871 : : * look at it, it can't be changing *to* a value we care about, only
4872 : : * *away* from such a value. So false negatives are impossible, and
4873 : : * false positives are safe because we'll recheck after getting the
4874 : : * buffer lock.
4875 : : *
4876 : : * We could check forkNum and blockNum as well as the rlocator, but
4877 : : * the incremental win from doing so seems small.
4878 : : */
4879 [ + + ]: 3348352 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4880 : 3339999 : continue;
4881 : :
4882 : 8353 : LockBufHdr(bufHdr);
4883 : :
4884 [ + + ]: 21151 : for (j = 0; j < nforks; j++)
4885 : : {
4886 [ + - ]: 14926 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4887 [ + + ]: 14926 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4888 [ + + ]: 8255 : bufHdr->tag.blockNum >= firstDelBlock[j])
4889 : : {
4890 : 2128 : InvalidateBuffer(bufHdr); /* releases spinlock */
4891 : 2128 : break;
4892 : : }
4893 : : }
4894 [ + + ]: 8353 : if (j >= nforks)
4895 : 6225 : UnlockBufHdr(bufHdr);
4896 : : }
4897 : : }
4898 : :
4899 : : /* ---------------------------------------------------------------------
4900 : : * DropRelationsAllBuffers
4901 : : *
4902 : : * This function removes from the buffer pool all the pages of all
4903 : : * forks of the specified relations. It's equivalent to calling
4904 : : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4905 : : * --------------------------------------------------------------------
4906 : : */
4907 : : void
4908 : 18297 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4909 : : {
4910 : : int i;
4911 : 18297 : int n = 0;
4912 : : SMgrRelation *rels;
4913 : : BlockNumber (*block)[MAX_FORKNUM + 1];
4914 : 18297 : uint64 nBlocksToInvalidate = 0;
4915 : : RelFileLocator *locators;
4916 : 18297 : bool cached = true;
4917 : : bool use_bsearch;
4918 : :
4919 [ - + ]: 18297 : if (nlocators == 0)
4920 : 0 : return;
4921 : :
4922 : 18297 : rels = palloc_array(SMgrRelation, nlocators); /* non-local relations */
4923 : :
4924 : : /* If it's a local relation, it's localbuf.c's problem. */
4925 [ + + ]: 80304 : for (i = 0; i < nlocators; i++)
4926 : : {
4927 [ + + ]: 62007 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4928 : : {
4929 [ + + ]: 4475 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4930 : 4473 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4931 : : }
4932 : : else
4933 : 57532 : rels[n++] = smgr_reln[i];
4934 : : }
4935 : :
4936 : : /*
4937 : : * If there are no non-local relations, then we're done. Release the
4938 : : * memory and return.
4939 : : */
4940 [ + + ]: 18297 : if (n == 0)
4941 : : {
4942 : 1197 : pfree(rels);
4943 : 1197 : return;
4944 : : }
4945 : :
4946 : : /*
4947 : : * This is used to remember the number of blocks for all the relations
4948 : : * forks.
4949 : : */
4950 : : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4951 : 17100 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4952 : :
4953 : : /*
4954 : : * We can avoid scanning the entire buffer pool if we know the exact size
4955 : : * of each of the given relation forks. See DropRelationBuffers.
4956 : : */
4957 [ + + + + ]: 35647 : for (i = 0; i < n && cached; i++)
4958 : : {
4959 [ + + ]: 27878 : for (int j = 0; j <= MAX_FORKNUM; j++)
4960 : : {
4961 : : /* Get the number of blocks for a relation's fork. */
4962 : 25563 : block[i][j] = smgrnblocks_cached(rels[i], j);
4963 : :
4964 : : /* We need to only consider the relation forks that exists. */
4965 [ + + ]: 25563 : if (block[i][j] == InvalidBlockNumber)
4966 : : {
4967 [ + + ]: 23041 : if (!smgrexists(rels[i], j))
4968 : 6809 : continue;
4969 : 16232 : cached = false;
4970 : 16232 : break;
4971 : : }
4972 : :
4973 : : /* calculate the total number of blocks to be invalidated */
4974 : 2522 : nBlocksToInvalidate += block[i][j];
4975 : : }
4976 : : }
4977 : :
4978 : : /*
4979 : : * We apply the optimization iff the total number of blocks to invalidate
4980 : : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4981 : : */
4982 [ + + + + ]: 17100 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4983 : : {
4984 [ + + ]: 1447 : for (i = 0; i < n; i++)
4985 : : {
4986 [ + + ]: 4005 : for (int j = 0; j <= MAX_FORKNUM; j++)
4987 : : {
4988 : : /* ignore relation forks that doesn't exist */
4989 [ + + ]: 3204 : if (!BlockNumberIsValid(block[i][j]))
4990 : 2392 : continue;
4991 : :
4992 : : /* drop all the buffers for a particular relation fork */
4993 : 812 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4994 : 812 : j, block[i][j], 0);
4995 : : }
4996 : : }
4997 : :
4998 : 646 : pfree(block);
4999 : 646 : pfree(rels);
5000 : 646 : return;
5001 : : }
5002 : :
5003 : 16454 : pfree(block);
5004 : 16454 : locators = palloc_array(RelFileLocator, n); /* non-local relations */
5005 [ + + ]: 73185 : for (i = 0; i < n; i++)
5006 : 56731 : locators[i] = rels[i]->smgr_rlocator.locator;
5007 : :
5008 : : /*
5009 : : * For low number of relations to drop just use a simple walk through, to
5010 : : * save the bsearch overhead. The threshold to use is rather a guess than
5011 : : * an exactly determined value, as it depends on many factors (CPU and RAM
5012 : : * speeds, amount of shared buffers etc.).
5013 : : */
5014 : 16454 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
5015 : :
5016 : : /* sort the list of rlocators if necessary */
5017 [ + + ]: 16454 : if (use_bsearch)
5018 : 218 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
5019 : :
5020 [ + + ]: 192008902 : for (i = 0; i < NBuffers; i++)
5021 : : {
5022 : 191992448 : RelFileLocator *rlocator = NULL;
5023 : 191992448 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5024 : :
5025 : : /*
5026 : : * As in DropRelationBuffers, an unlocked precheck should be safe and
5027 : : * saves some cycles.
5028 : : */
5029 : :
5030 [ + + ]: 191992448 : if (!use_bsearch)
5031 : : {
5032 : : int j;
5033 : :
5034 [ + + ]: 763598028 : for (j = 0; j < n; j++)
5035 : : {
5036 [ + + ]: 574210967 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
5037 : : {
5038 : 106571 : rlocator = &locators[j];
5039 : 106571 : break;
5040 : : }
5041 : : }
5042 : : }
5043 : : else
5044 : : {
5045 : : RelFileLocator locator;
5046 : :
5047 : 2498816 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
5048 : 2498816 : rlocator = bsearch(&locator,
5049 : : locators, n, sizeof(RelFileLocator),
5050 : : rlocator_comparator);
5051 : : }
5052 : :
5053 : : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5054 [ + + ]: 191992448 : if (rlocator == NULL)
5055 : 191884059 : continue;
5056 : :
5057 : 108389 : LockBufHdr(bufHdr);
5058 [ + - ]: 108389 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
5059 : 108389 : InvalidateBuffer(bufHdr); /* releases spinlock */
5060 : : else
5061 : 0 : UnlockBufHdr(bufHdr);
5062 : : }
5063 : :
5064 : 16454 : pfree(locators);
5065 : 16454 : pfree(rels);
5066 : : }
5067 : :
5068 : : /* ---------------------------------------------------------------------
5069 : : * FindAndDropRelationBuffers
5070 : : *
5071 : : * This function performs look up in BufMapping table and removes from the
5072 : : * buffer pool all the pages of the specified relation fork that has block
5073 : : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
5074 : : * pages are removed.)
5075 : : * --------------------------------------------------------------------
5076 : : */
5077 : : static void
5078 : 880 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
5079 : : BlockNumber nForkBlock,
5080 : : BlockNumber firstDelBlock)
5081 : : {
5082 : : BlockNumber curBlock;
5083 : :
5084 [ + + ]: 2111 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
5085 : : {
5086 : : uint32 bufHash; /* hash value for tag */
5087 : : BufferTag bufTag; /* identity of requested block */
5088 : : LWLock *bufPartitionLock; /* buffer partition lock for it */
5089 : : int buf_id;
5090 : : BufferDesc *bufHdr;
5091 : :
5092 : : /* create a tag so we can lookup the buffer */
5093 : 1231 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
5094 : :
5095 : : /* determine its hash code and partition lock ID */
5096 : 1231 : bufHash = BufTableHashCode(&bufTag);
5097 : 1231 : bufPartitionLock = BufMappingPartitionLock(bufHash);
5098 : :
5099 : : /* Check that it is in the buffer pool. If not, do nothing. */
5100 : 1231 : LWLockAcquire(bufPartitionLock, LW_SHARED);
5101 : 1231 : buf_id = BufTableLookup(&bufTag, bufHash);
5102 : 1231 : LWLockRelease(bufPartitionLock);
5103 : :
5104 [ + + ]: 1231 : if (buf_id < 0)
5105 : 133 : continue;
5106 : :
5107 : 1098 : bufHdr = GetBufferDescriptor(buf_id);
5108 : :
5109 : : /*
5110 : : * We need to lock the buffer header and recheck if the buffer is
5111 : : * still associated with the same block because the buffer could be
5112 : : * evicted by some other backend loading blocks for a different
5113 : : * relation after we release lock on the BufMapping table.
5114 : : */
5115 : 1098 : LockBufHdr(bufHdr);
5116 : :
5117 [ + - + - ]: 2196 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
5118 : 1098 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
5119 [ + - ]: 1098 : bufHdr->tag.blockNum >= firstDelBlock)
5120 : 1098 : InvalidateBuffer(bufHdr); /* releases spinlock */
5121 : : else
5122 : 0 : UnlockBufHdr(bufHdr);
5123 : : }
5124 : 880 : }
5125 : :
5126 : : /* ---------------------------------------------------------------------
5127 : : * DropDatabaseBuffers
5128 : : *
5129 : : * This function removes all the buffers in the buffer cache for a
5130 : : * particular database. Dirty pages are simply dropped, without
5131 : : * bothering to write them out first. This is used when we destroy a
5132 : : * database, to avoid trying to flush data to disk when the directory
5133 : : * tree no longer exists. Implementation is pretty similar to
5134 : : * DropRelationBuffers() which is for destroying just one relation.
5135 : : * --------------------------------------------------------------------
5136 : : */
5137 : : void
5138 : 84 : DropDatabaseBuffers(Oid dbid)
5139 : : {
5140 : : int i;
5141 : :
5142 : : /*
5143 : : * We needn't consider local buffers, since by assumption the target
5144 : : * database isn't our own.
5145 : : */
5146 : :
5147 [ + + ]: 644820 : for (i = 0; i < NBuffers; i++)
5148 : : {
5149 : 644736 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5150 : :
5151 : : /*
5152 : : * As in DropRelationBuffers, an unlocked precheck should be safe and
5153 : : * saves some cycles.
5154 : : */
5155 [ + + ]: 644736 : if (bufHdr->tag.dbOid != dbid)
5156 : 628908 : continue;
5157 : :
5158 : 15828 : LockBufHdr(bufHdr);
5159 [ + - ]: 15828 : if (bufHdr->tag.dbOid == dbid)
5160 : 15828 : InvalidateBuffer(bufHdr); /* releases spinlock */
5161 : : else
5162 : 0 : UnlockBufHdr(bufHdr);
5163 : : }
5164 : 84 : }
5165 : :
5166 : : /* ---------------------------------------------------------------------
5167 : : * FlushRelationBuffers
5168 : : *
5169 : : * This function writes all dirty pages of a relation out to disk
5170 : : * (or more accurately, out to kernel disk buffers), ensuring that the
5171 : : * kernel has an up-to-date view of the relation.
5172 : : *
5173 : : * Generally, the caller should be holding AccessExclusiveLock on the
5174 : : * target relation to ensure that no other backend is busy dirtying
5175 : : * more blocks of the relation; the effects can't be expected to last
5176 : : * after the lock is released.
5177 : : *
5178 : : * XXX currently it sequentially searches the buffer pool, should be
5179 : : * changed to more clever ways of searching. This routine is not
5180 : : * used in any performance-critical code paths, so it's not worth
5181 : : * adding additional overhead to normal paths to make it go faster.
5182 : : * --------------------------------------------------------------------
5183 : : */
5184 : : void
5185 : 186 : FlushRelationBuffers(Relation rel)
5186 : : {
5187 : : int i;
5188 : : BufferDesc *bufHdr;
5189 : 186 : SMgrRelation srel = RelationGetSmgr(rel);
5190 : :
5191 [ + + ]: 186 : if (RelationUsesLocalBuffers(rel))
5192 : : {
5193 [ + + ]: 1212 : for (i = 0; i < NLocBuffer; i++)
5194 : : {
5195 : : uint64 buf_state;
5196 : :
5197 : 1200 : bufHdr = GetLocalBufferDescriptor(i);
5198 [ + + ]: 1200 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5199 [ + + ]: 400 : ((buf_state = pg_atomic_read_u64(&bufHdr->state)) &
5200 : : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5201 : : {
5202 : : ErrorContextCallback errcallback;
5203 : :
5204 : : /* Setup error traceback support for ereport() */
5205 : 392 : errcallback.callback = local_buffer_write_error_callback;
5206 : 392 : errcallback.arg = bufHdr;
5207 : 392 : errcallback.previous = error_context_stack;
5208 : 392 : error_context_stack = &errcallback;
5209 : :
5210 : : /* Make sure we can handle the pin */
5211 : 392 : ReservePrivateRefCountEntry();
5212 : 392 : ResourceOwnerEnlarge(CurrentResourceOwner);
5213 : :
5214 : : /*
5215 : : * Pin/unpin mostly to make valgrind work, but it also seems
5216 : : * like the right thing to do.
5217 : : */
5218 : 392 : PinLocalBuffer(bufHdr, false);
5219 : :
5220 : :
5221 : 392 : FlushLocalBuffer(bufHdr, srel);
5222 : :
5223 : 392 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
5224 : :
5225 : : /* Pop the error context stack */
5226 : 392 : error_context_stack = errcallback.previous;
5227 : : }
5228 : : }
5229 : :
5230 : 12 : return;
5231 : : }
5232 : :
5233 [ + + ]: 2021934 : for (i = 0; i < NBuffers; i++)
5234 : : {
5235 : : uint64 buf_state;
5236 : :
5237 : 2021760 : bufHdr = GetBufferDescriptor(i);
5238 : :
5239 : : /*
5240 : : * As in DropRelationBuffers, an unlocked precheck should be safe and
5241 : : * saves some cycles.
5242 : : */
5243 [ + + ]: 2021760 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5244 : 2021504 : continue;
5245 : :
5246 : : /* Make sure we can handle the pin */
5247 : 256 : ReservePrivateRefCountEntry();
5248 : 256 : ResourceOwnerEnlarge(CurrentResourceOwner);
5249 : :
5250 : 256 : buf_state = LockBufHdr(bufHdr);
5251 [ + - ]: 256 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5252 [ + + ]: 256 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5253 : : {
5254 : 208 : PinBuffer_Locked(bufHdr);
5255 : 208 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5256 : 208 : UnpinBuffer(bufHdr);
5257 : : }
5258 : : else
5259 : 48 : UnlockBufHdr(bufHdr);
5260 : : }
5261 : : }
5262 : :
5263 : : /* ---------------------------------------------------------------------
5264 : : * FlushRelationsAllBuffers
5265 : : *
5266 : : * This function flushes out of the buffer pool all the pages of all
5267 : : * forks of the specified smgr relations. It's equivalent to calling
5268 : : * FlushRelationBuffers once per relation. The relations are assumed not
5269 : : * to use local buffers.
5270 : : * --------------------------------------------------------------------
5271 : : */
5272 : : void
5273 : 10 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5274 : : {
5275 : : int i;
5276 : : SMgrSortArray *srels;
5277 : : bool use_bsearch;
5278 : :
5279 [ - + ]: 10 : if (nrels == 0)
5280 : 0 : return;
5281 : :
5282 : : /* fill-in array for qsort */
5283 : 10 : srels = palloc_array(SMgrSortArray, nrels);
5284 : :
5285 [ + + ]: 31 : for (i = 0; i < nrels; i++)
5286 : : {
5287 : : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5288 : :
5289 : 21 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5290 : 21 : srels[i].srel = smgrs[i];
5291 : : }
5292 : :
5293 : : /*
5294 : : * Save the bsearch overhead for low number of relations to sync. See
5295 : : * DropRelationsAllBuffers for details.
5296 : : */
5297 : 10 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5298 : :
5299 : : /* sort the list of SMgrRelations if necessary */
5300 [ - + ]: 10 : if (use_bsearch)
5301 : 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5302 : :
5303 [ + + ]: 163850 : for (i = 0; i < NBuffers; i++)
5304 : : {
5305 : 163840 : SMgrSortArray *srelent = NULL;
5306 : 163840 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5307 : : uint64 buf_state;
5308 : :
5309 : : /*
5310 : : * As in DropRelationBuffers, an unlocked precheck should be safe and
5311 : : * saves some cycles.
5312 : : */
5313 : :
5314 [ + - ]: 163840 : if (!use_bsearch)
5315 : : {
5316 : : int j;
5317 : :
5318 [ + + ]: 505265 : for (j = 0; j < nrels; j++)
5319 : : {
5320 [ + + ]: 343857 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5321 : : {
5322 : 2432 : srelent = &srels[j];
5323 : 2432 : break;
5324 : : }
5325 : : }
5326 : : }
5327 : : else
5328 : : {
5329 : : RelFileLocator rlocator;
5330 : :
5331 : 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5332 : 0 : srelent = bsearch(&rlocator,
5333 : : srels, nrels, sizeof(SMgrSortArray),
5334 : : rlocator_comparator);
5335 : : }
5336 : :
5337 : : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5338 [ + + ]: 163840 : if (srelent == NULL)
5339 : 161408 : continue;
5340 : :
5341 : : /* Make sure we can handle the pin */
5342 : 2432 : ReservePrivateRefCountEntry();
5343 : 2432 : ResourceOwnerEnlarge(CurrentResourceOwner);
5344 : :
5345 : 2432 : buf_state = LockBufHdr(bufHdr);
5346 [ + - ]: 2432 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5347 [ + + ]: 2432 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5348 : : {
5349 : 2391 : PinBuffer_Locked(bufHdr);
5350 : 2391 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5351 : 2391 : UnpinBuffer(bufHdr);
5352 : : }
5353 : : else
5354 : 41 : UnlockBufHdr(bufHdr);
5355 : : }
5356 : :
5357 : 10 : pfree(srels);
5358 : : }
5359 : :
5360 : : /* ---------------------------------------------------------------------
5361 : : * RelationCopyStorageUsingBuffer
5362 : : *
5363 : : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5364 : : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5365 : : *
5366 : : * Refer comments atop CreateAndCopyRelationData() for details about
5367 : : * 'permanent' parameter.
5368 : : * --------------------------------------------------------------------
5369 : : */
5370 : : static void
5371 : 84620 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5372 : : RelFileLocator dstlocator,
5373 : : ForkNumber forkNum, bool permanent)
5374 : : {
5375 : : Buffer srcBuf;
5376 : : Buffer dstBuf;
5377 : : Page srcPage;
5378 : : Page dstPage;
5379 : : bool use_wal;
5380 : : BlockNumber nblocks;
5381 : : BlockNumber blkno;
5382 : : PGIOAlignedBlock buf;
5383 : : BufferAccessStrategy bstrategy_src;
5384 : : BufferAccessStrategy bstrategy_dst;
5385 : : BlockRangeReadStreamPrivate p;
5386 : : ReadStream *src_stream;
5387 : : SMgrRelation src_smgr;
5388 : :
5389 : : /*
5390 : : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5391 : : * can skip it when copying any fork of an unlogged relation other than
5392 : : * the init fork.
5393 : : */
5394 [ + + - + : 84620 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
- - ]
5395 : :
5396 : : /* Get number of blocks in the source relation. */
5397 : 84620 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5398 : : forkNum);
5399 : :
5400 : : /* Nothing to copy; just return. */
5401 [ + + ]: 84620 : if (nblocks == 0)
5402 : 15697 : return;
5403 : :
5404 : : /*
5405 : : * Bulk extend the destination relation of the same size as the source
5406 : : * relation before starting to copy block by block.
5407 : : */
5408 : 68923 : memset(buf.data, 0, BLCKSZ);
5409 : 68923 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5410 : : buf.data, true);
5411 : :
5412 : : /* This is a bulk operation, so use buffer access strategies. */
5413 : 68923 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5414 : 68923 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5415 : :
5416 : : /* Initialize streaming read */
5417 : 68923 : p.current_blocknum = 0;
5418 : 68923 : p.last_exclusive = nblocks;
5419 : 68923 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5420 : :
5421 : : /*
5422 : : * It is safe to use batchmode as block_range_read_stream_cb takes no
5423 : : * locks.
5424 : : */
5425 [ + - ]: 68923 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5426 : : READ_STREAM_USE_BATCHING,
5427 : : bstrategy_src,
5428 : : src_smgr,
5429 : : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5430 : : forkNum,
5431 : : block_range_read_stream_cb,
5432 : : &p,
5433 : : 0);
5434 : :
5435 : : /* Iterate over each block of the source relation file. */
5436 [ + + ]: 326269 : for (blkno = 0; blkno < nblocks; blkno++)
5437 : : {
5438 [ - + ]: 257348 : CHECK_FOR_INTERRUPTS();
5439 : :
5440 : : /* Read block from source relation. */
5441 : 257348 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5442 : 257346 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5443 : 257346 : srcPage = BufferGetPage(srcBuf);
5444 : :
5445 : 257346 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5446 : : BufferGetBlockNumber(srcBuf),
5447 : : RBM_ZERO_AND_LOCK, bstrategy_dst,
5448 : : permanent);
5449 : 257346 : dstPage = BufferGetPage(dstBuf);
5450 : :
5451 : 257346 : START_CRIT_SECTION();
5452 : :
5453 : : /* Copy page data from the source to the destination. */
5454 : 257346 : memcpy(dstPage, srcPage, BLCKSZ);
5455 : 257346 : MarkBufferDirty(dstBuf);
5456 : :
5457 : : /* WAL-log the copied page. */
5458 [ + + ]: 257346 : if (use_wal)
5459 : 149010 : log_newpage_buffer(dstBuf, true);
5460 : :
5461 : 257346 : END_CRIT_SECTION();
5462 : :
5463 : 257346 : UnlockReleaseBuffer(dstBuf);
5464 : 257346 : UnlockReleaseBuffer(srcBuf);
5465 : : }
5466 : : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5467 : 68921 : read_stream_end(src_stream);
5468 : :
5469 : 68921 : FreeAccessStrategy(bstrategy_src);
5470 : 68921 : FreeAccessStrategy(bstrategy_dst);
5471 : : }
5472 : :
5473 : : /* ---------------------------------------------------------------------
5474 : : * CreateAndCopyRelationData
5475 : : *
5476 : : * Create destination relation storage and copy all forks from the
5477 : : * source relation to the destination.
5478 : : *
5479 : : * Pass permanent as true for permanent relations and false for
5480 : : * unlogged relations. Currently this API is not supported for
5481 : : * temporary relations.
5482 : : * --------------------------------------------------------------------
5483 : : */
5484 : : void
5485 : 64932 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5486 : : RelFileLocator dst_rlocator, bool permanent)
5487 : : {
5488 : : char relpersistence;
5489 : : SMgrRelation src_rel;
5490 : : SMgrRelation dst_rel;
5491 : :
5492 : : /* Set the relpersistence. */
5493 [ + - ]: 64932 : relpersistence = permanent ?
5494 : : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5495 : :
5496 : 64932 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5497 : 64932 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5498 : :
5499 : : /*
5500 : : * Create and copy all forks of the relation. During create database we
5501 : : * have a separate cleanup mechanism which deletes complete database
5502 : : * directory. Therefore, each individual relation doesn't need to be
5503 : : * registered for cleanup.
5504 : : */
5505 : 64932 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5506 : :
5507 : : /* copy main fork. */
5508 : 64932 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5509 : : permanent);
5510 : :
5511 : : /* copy those extra forks that exist */
5512 : 64930 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5513 [ + + ]: 259720 : forkNum <= MAX_FORKNUM; forkNum++)
5514 : : {
5515 [ + + ]: 194790 : if (smgrexists(src_rel, forkNum))
5516 : : {
5517 : 19688 : smgrcreate(dst_rel, forkNum, false);
5518 : :
5519 : : /*
5520 : : * WAL log creation if the relation is persistent, or this is the
5521 : : * init fork of an unlogged relation.
5522 : : */
5523 [ - + - - ]: 19688 : if (permanent || forkNum == INIT_FORKNUM)
5524 : 19688 : log_smgrcreate(&dst_rlocator, forkNum);
5525 : :
5526 : : /* Copy a fork's data, block by block. */
5527 : 19688 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5528 : : permanent);
5529 : : }
5530 : : }
5531 : 64930 : }
5532 : :
5533 : : /* ---------------------------------------------------------------------
5534 : : * FlushDatabaseBuffers
5535 : : *
5536 : : * This function writes all dirty pages of a database out to disk
5537 : : * (or more accurately, out to kernel disk buffers), ensuring that the
5538 : : * kernel has an up-to-date view of the database.
5539 : : *
5540 : : * Generally, the caller should be holding an appropriate lock to ensure
5541 : : * no other backend is active in the target database; otherwise more
5542 : : * pages could get dirtied.
5543 : : *
5544 : : * Note we don't worry about flushing any pages of temporary relations.
5545 : : * It's assumed these wouldn't be interesting.
5546 : : * --------------------------------------------------------------------
5547 : : */
5548 : : void
5549 : 5 : FlushDatabaseBuffers(Oid dbid)
5550 : : {
5551 : : int i;
5552 : : BufferDesc *bufHdr;
5553 : :
5554 [ + + ]: 645 : for (i = 0; i < NBuffers; i++)
5555 : : {
5556 : : uint64 buf_state;
5557 : :
5558 : 640 : bufHdr = GetBufferDescriptor(i);
5559 : :
5560 : : /*
5561 : : * As in DropRelationBuffers, an unlocked precheck should be safe and
5562 : : * saves some cycles.
5563 : : */
5564 [ + + ]: 640 : if (bufHdr->tag.dbOid != dbid)
5565 : 477 : continue;
5566 : :
5567 : : /* Make sure we can handle the pin */
5568 : 163 : ReservePrivateRefCountEntry();
5569 : 163 : ResourceOwnerEnlarge(CurrentResourceOwner);
5570 : :
5571 : 163 : buf_state = LockBufHdr(bufHdr);
5572 [ + - ]: 163 : if (bufHdr->tag.dbOid == dbid &&
5573 [ + + ]: 163 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5574 : : {
5575 : 27 : PinBuffer_Locked(bufHdr);
5576 : 27 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5577 : 27 : UnpinBuffer(bufHdr);
5578 : : }
5579 : : else
5580 : 136 : UnlockBufHdr(bufHdr);
5581 : : }
5582 : 5 : }
5583 : :
5584 : : /*
5585 : : * Flush a previously, share-exclusively or exclusively, locked and pinned
5586 : : * buffer to the OS.
5587 : : */
5588 : : void
5589 : 112 : FlushOneBuffer(Buffer buffer)
5590 : : {
5591 : : BufferDesc *bufHdr;
5592 : :
5593 : : /* currently not needed, but no fundamental reason not to support */
5594 : : Assert(!BufferIsLocal(buffer));
5595 : :
5596 : : Assert(BufferIsPinned(buffer));
5597 : :
5598 : 112 : bufHdr = GetBufferDescriptor(buffer - 1);
5599 : :
5600 : : Assert(BufferIsLockedByMe(buffer));
5601 : :
5602 : 112 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5603 : 112 : }
5604 : :
5605 : : /*
5606 : : * ReleaseBuffer -- release the pin on a buffer
5607 : : */
5608 : : void
5609 : 50025923 : ReleaseBuffer(Buffer buffer)
5610 : : {
5611 [ - + ]: 50025923 : if (!BufferIsValid(buffer))
5612 [ # # ]: 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5613 : :
5614 [ + + ]: 50025923 : if (BufferIsLocal(buffer))
5615 : 752280 : UnpinLocalBuffer(buffer);
5616 : : else
5617 : 49273643 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5618 : 50025923 : }
5619 : :
5620 : : /*
5621 : : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5622 : : *
5623 : : * This is just a, more efficient, shorthand for a common combination.
5624 : : */
5625 : : void
5626 : 51118579 : UnlockReleaseBuffer(Buffer buffer)
5627 : : {
5628 : : int mode;
5629 : : BufferDesc *buf;
5630 : : PrivateRefCountEntry *ref;
5631 : : uint64 sub;
5632 : : uint64 lockstate;
5633 : :
5634 : : Assert(BufferIsPinned(buffer));
5635 : :
5636 [ + + ]: 51118579 : if (BufferIsLocal(buffer))
5637 : : {
5638 : 1381960 : UnpinLocalBuffer(buffer);
5639 : 1381960 : return;
5640 : : }
5641 : :
5642 : 49736619 : ResourceOwnerForgetBuffer(CurrentResourceOwner, buffer);
5643 : :
5644 : 49736619 : buf = GetBufferDescriptor(buffer - 1);
5645 : :
5646 : 49736619 : mode = BufferLockDisownInternal(buffer, buf);
5647 : :
5648 : : /* compute state modification for lock release */
5649 : 49736619 : sub = BufferLockReleaseSub(mode);
5650 : :
5651 : : /* compute state modification for pin release */
5652 : 49736619 : ref = GetPrivateRefCountEntry(buffer, false);
5653 : : Assert(ref != NULL);
5654 : : Assert(ref->data.refcount > 0);
5655 : 49736619 : ref->data.refcount--;
5656 : :
5657 : : /* no more backend local pins, reduce shared pin count */
5658 [ + + ]: 49736619 : if (likely(ref->data.refcount == 0))
5659 : : {
5660 : : /* See comment in UnpinBufferNoOwner() */
5661 : : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
5662 : :
5663 : 47355974 : sub |= BUF_REFCOUNT_ONE;
5664 : 47355974 : ForgetPrivateRefCountEntry(ref);
5665 : : }
5666 : :
5667 : : /* perform the lock and pin release in one atomic op */
5668 : 49736619 : lockstate = pg_atomic_sub_fetch_u64(&buf->state, sub);
5669 : :
5670 : : /* wake up waiters for the lock */
5671 : 49736619 : BufferLockProcessRelease(buf, mode, lockstate);
5672 : :
5673 : : /* wake up waiter for the pin release */
5674 [ + + ]: 49736619 : if (lockstate & BM_PIN_COUNT_WAITER)
5675 : 1 : WakePinCountWaiter(buf);
5676 : :
5677 : : /*
5678 : : * Now okay to allow cancel/die interrupts again, which were held when the
5679 : : * lock was acquired.
5680 : : */
5681 : 49736619 : RESUME_INTERRUPTS();
5682 : : }
5683 : :
5684 : : /*
5685 : : * IncrBufferRefCount
5686 : : * Increment the pin count on a buffer that we have *already* pinned
5687 : : * at least once.
5688 : : *
5689 : : * This function cannot be used on a buffer we do not have pinned,
5690 : : * because it doesn't change the shared buffer state.
5691 : : */
5692 : : void
5693 : 15043375 : IncrBufferRefCount(Buffer buffer)
5694 : : {
5695 : : Assert(BufferIsPinned(buffer));
5696 : 15043375 : ResourceOwnerEnlarge(CurrentResourceOwner);
5697 [ + + ]: 15043375 : if (BufferIsLocal(buffer))
5698 : 470018 : LocalRefCount[-buffer - 1]++;
5699 : : else
5700 : : {
5701 : : PrivateRefCountEntry *ref;
5702 : :
5703 : 14573357 : ref = GetPrivateRefCountEntry(buffer, true);
5704 : : Assert(ref != NULL);
5705 : 14573357 : ref->data.refcount++;
5706 : : }
5707 : 15043375 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5708 : 15043375 : }
5709 : :
5710 : : /*
5711 : : * Shared-buffer only helper for MarkBufferDirtyHint() and
5712 : : * BufferSetHintBits16().
5713 : : *
5714 : : * This is separated out because it turns out that the repeated checks for
5715 : : * local buffers, repeated GetBufferDescriptor() and repeated reading of the
5716 : : * buffer's state sufficiently hurts the performance of BufferSetHintBits16().
5717 : : */
5718 : : static inline void
5719 : 15190100 : MarkSharedBufferDirtyHint(Buffer buffer, BufferDesc *bufHdr, uint64 lockstate,
5720 : : bool buffer_std)
5721 : : {
5722 : 15190100 : Page page = BufferGetPage(buffer);
5723 : :
5724 : : Assert(GetPrivateRefCount(buffer) > 0);
5725 : :
5726 : : /* here, either share-exclusive or exclusive lock is OK */
5727 : : Assert(BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_EXCLUSIVE) ||
5728 : : BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_SHARE_EXCLUSIVE));
5729 : :
5730 : : /*
5731 : : * This routine might get called many times on the same page, if we are
5732 : : * making the first scan after commit of an xact that added/deleted many
5733 : : * tuples. So, be as quick as we can if the buffer is already dirty.
5734 : : *
5735 : : * As we are holding (at least) a share-exclusive lock, nobody could have
5736 : : * cleaned or dirtied the page concurrently, so we can just rely on the
5737 : : * previously fetched value here without any danger of races.
5738 : : */
5739 [ + + ]: 15190100 : if (unlikely(!(lockstate & BM_DIRTY)))
5740 : : {
5741 : 423199 : XLogRecPtr lsn = InvalidXLogRecPtr;
5742 : 423199 : bool wal_log = false;
5743 : : uint64 buf_state;
5744 : :
5745 : : /*
5746 : : * If we need to protect hint bit updates from torn writes, WAL-log a
5747 : : * full page image of the page. This full page image is only necessary
5748 : : * if the hint bit update is the first change to the page since the
5749 : : * last checkpoint.
5750 : : *
5751 : : * We don't check full_page_writes here because that logic is included
5752 : : * when we call XLogInsert() since the value changes dynamically.
5753 : : */
5754 [ + + + + : 423199 : if (XLogHintBitIsNeeded() && (lockstate & BM_PERMANENT))
+ + ]
5755 : : {
5756 : : /*
5757 : : * If we must not write WAL, due to a relfilelocator-specific
5758 : : * condition or being in recovery, don't dirty the page. We can
5759 : : * set the hint, just not dirty the page as a result so the hint
5760 : : * is lost when we evict the page or shutdown.
5761 : : *
5762 : : * See src/backend/storage/page/README for longer discussion.
5763 : : */
5764 [ + + + + ]: 507182 : if (RecoveryInProgress() ||
5765 : 85976 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5766 : 336478 : return;
5767 : :
5768 : 84728 : wal_log = true;
5769 : : }
5770 : :
5771 : : /*
5772 : : * We must mark the page dirty before we emit the WAL record, as per
5773 : : * the usual rules, to ensure that BufferSync()/SyncOneBuffer() try to
5774 : : * flush the buffer, even if we haven't inserted the WAL record yet.
5775 : : * As we hold at least a share-exclusive lock, checkpoints will wait
5776 : : * for this backend to be done with the buffer before continuing. If
5777 : : * we did it the other way round, a checkpoint could start between
5778 : : * writing the WAL record and marking the buffer dirty.
5779 : : */
5780 : 86721 : buf_state = LockBufHdr(bufHdr);
5781 : :
5782 : : /*
5783 : : * It should not be possible for the buffer to already be dirty, see
5784 : : * comment above.
5785 : : */
5786 : : Assert(!(buf_state & BM_DIRTY));
5787 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5788 : 86721 : UnlockBufHdrExt(bufHdr, buf_state,
5789 : : BM_DIRTY,
5790 : : 0, 0);
5791 : :
5792 : : /*
5793 : : * If the block is already dirty because we either made a change or
5794 : : * set a hint already, then we don't need to write a full page image.
5795 : : * Note that aggressive cleaning of blocks dirtied by hint bit setting
5796 : : * would increase the call rate. Bulk setting of hint bits would
5797 : : * reduce the call rate...
5798 : : */
5799 [ + + ]: 86721 : if (wal_log)
5800 : 84728 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5801 : :
5802 [ + + ]: 86721 : if (XLogRecPtrIsValid(lsn))
5803 : : {
5804 : : /*
5805 : : * Set the page LSN if we wrote a backup block. To allow backends
5806 : : * that only hold a share lock on the buffer to read the LSN in a
5807 : : * tear-free manner, we set the page LSN while holding the buffer
5808 : : * header lock. This allows any reader of an LSN who holds only a
5809 : : * share lock to also obtain a buffer header lock before using
5810 : : * PageGetLSN() to read the LSN in a tear free way. This is done
5811 : : * in BufferGetLSNAtomic().
5812 : : *
5813 : : * If checksums are enabled, you might think we should reset the
5814 : : * checksum here. That will happen when the page is written
5815 : : * sometime later in this checkpoint cycle.
5816 : : */
5817 : 58422 : buf_state = LockBufHdr(bufHdr);
5818 : 58422 : PageSetLSN(page, lsn);
5819 : 58422 : UnlockBufHdr(bufHdr);
5820 : : }
5821 : :
5822 : 86721 : pgBufferUsage.shared_blks_dirtied++;
5823 [ + + ]: 86721 : if (VacuumCostActive)
5824 : 1447 : VacuumCostBalance += VacuumCostPageDirty;
5825 : : }
5826 : : }
5827 : :
5828 : : /*
5829 : : * MarkBufferDirtyHint
5830 : : *
5831 : : * Mark a buffer dirty for non-critical changes.
5832 : : *
5833 : : * This is essentially the same as MarkBufferDirty, except:
5834 : : *
5835 : : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5836 : : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5837 : : * 2. The caller might have only a share-exclusive-lock instead of an
5838 : : * exclusive-lock on the buffer's content lock.
5839 : : * 3. This function does not guarantee that the buffer is always marked dirty
5840 : : * (it e.g. can't always on a hot standby), so it cannot be used for
5841 : : * important changes.
5842 : : */
5843 : : inline void
5844 : 439934 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5845 : : {
5846 : : BufferDesc *bufHdr;
5847 : :
5848 [ - + ]: 439934 : if (!BufferIsValid(buffer))
5849 [ # # ]: 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5850 : :
5851 [ + + ]: 439934 : if (BufferIsLocal(buffer))
5852 : : {
5853 : 22052 : MarkLocalBufferDirty(buffer);
5854 : 22052 : return;
5855 : : }
5856 : :
5857 : 417882 : bufHdr = GetBufferDescriptor(buffer - 1);
5858 : :
5859 : 417882 : MarkSharedBufferDirtyHint(buffer, bufHdr,
5860 : 417882 : pg_atomic_read_u64(&bufHdr->state),
5861 : : buffer_std);
5862 : : }
5863 : :
5864 : : /*
5865 : : * Release buffer content locks for shared buffers.
5866 : : *
5867 : : * Used to clean up after errors.
5868 : : *
5869 : : * Currently, we can expect that resource owner cleanup, via
5870 : : * ResOwnerReleaseBuffer(), took care of releasing buffer content locks per
5871 : : * se; the only thing we need to deal with here is clearing any PIN_COUNT
5872 : : * request that was in progress.
5873 : : */
5874 : : void
5875 : 66181 : UnlockBuffers(void)
5876 : : {
5877 : 66181 : BufferDesc *buf = PinCountWaitBuf;
5878 : :
5879 [ - + ]: 66181 : if (buf)
5880 : : {
5881 : : uint64 buf_state;
5882 : 0 : uint64 unset_bits = 0;
5883 : :
5884 : 0 : buf_state = LockBufHdr(buf);
5885 : :
5886 : : /*
5887 : : * Don't complain if flag bit not set; it could have been reset but we
5888 : : * got a cancel/die interrupt before getting the signal.
5889 : : */
5890 [ # # ]: 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5891 [ # # ]: 0 : buf->wait_backend_pgprocno == MyProcNumber)
5892 : 0 : unset_bits = BM_PIN_COUNT_WAITER;
5893 : :
5894 : 0 : UnlockBufHdrExt(buf, buf_state,
5895 : : 0, unset_bits,
5896 : : 0);
5897 : :
5898 : 0 : PinCountWaitBuf = NULL;
5899 : : }
5900 : 66181 : }
5901 : :
5902 : : /*
5903 : : * Acquire the buffer content lock in the specified mode
5904 : : *
5905 : : * If the lock is not available, sleep until it is.
5906 : : *
5907 : : * Side effect: cancel/die interrupts are held off until lock release.
5908 : : *
5909 : : * This uses almost the same locking approach as lwlock.c's
5910 : : * LWLockAcquire(). See documentation at the top of lwlock.c for a more
5911 : : * detailed discussion.
5912 : : *
5913 : : * The reason that this, and most of the other BufferLock* functions, get both
5914 : : * the Buffer and BufferDesc* as parameters, is that looking up one from the
5915 : : * other repeatedly shows up noticeably in profiles.
5916 : : *
5917 : : * Callers should provide a constant for mode, for more efficient code
5918 : : * generation.
5919 : : */
5920 : : static inline void
5921 : 112116182 : BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
5922 : : {
5923 : : PrivateRefCountEntry *entry;
5924 : 112116182 : int extraWaits = 0;
5925 : :
5926 : : /*
5927 : : * Get reference to the refcount entry before we hold the lock, it seems
5928 : : * better to do before holding the lock.
5929 : : */
5930 : 112116182 : entry = GetPrivateRefCountEntry(buffer, true);
5931 : :
5932 : : /*
5933 : : * We better not already hold a lock on the buffer.
5934 : : */
5935 : : Assert(entry->data.lockmode == BUFFER_LOCK_UNLOCK);
5936 : :
5937 : : /*
5938 : : * Lock out cancel/die interrupts until we exit the code section protected
5939 : : * by the content lock. This ensures that interrupts will not interfere
5940 : : * with manipulations of data structures in shared memory.
5941 : : */
5942 : 112116182 : HOLD_INTERRUPTS();
5943 : :
5944 : : for (;;)
5945 : 22230 : {
5946 : 112138412 : uint32 wait_event = 0; /* initialized to avoid compiler warning */
5947 : : bool mustwait;
5948 : :
5949 : : /*
5950 : : * Try to grab the lock the first time, we're not in the waitqueue
5951 : : * yet/anymore.
5952 : : */
5953 : 112138412 : mustwait = BufferLockAttempt(buf_hdr, mode);
5954 : :
5955 [ + + ]: 112138412 : if (likely(!mustwait))
5956 : : {
5957 : 112114544 : break;
5958 : : }
5959 : :
5960 : : /*
5961 : : * Ok, at this point we couldn't grab the lock on the first try. We
5962 : : * cannot simply queue ourselves to the end of the list and wait to be
5963 : : * woken up because by now the lock could long have been released.
5964 : : * Instead add us to the queue and try to grab the lock again. If we
5965 : : * succeed we need to revert the queuing and be happy, otherwise we
5966 : : * recheck the lock. If we still couldn't grab it, we know that the
5967 : : * other locker will see our queue entries when releasing since they
5968 : : * existed before we checked for the lock.
5969 : : */
5970 : :
5971 : : /* add to the queue */
5972 : 23868 : BufferLockQueueSelf(buf_hdr, mode);
5973 : :
5974 : : /* we're now guaranteed to be woken up if necessary */
5975 : 23868 : mustwait = BufferLockAttempt(buf_hdr, mode);
5976 : :
5977 : : /* ok, grabbed the lock the second time round, need to undo queueing */
5978 [ + + ]: 23868 : if (!mustwait)
5979 : : {
5980 : 1638 : BufferLockDequeueSelf(buf_hdr);
5981 : 1638 : break;
5982 : : }
5983 : :
5984 [ + + + - ]: 22230 : switch (mode)
5985 : : {
5986 : 12752 : case BUFFER_LOCK_EXCLUSIVE:
5987 : 12752 : wait_event = WAIT_EVENT_BUFFER_EXCLUSIVE;
5988 : 12752 : break;
5989 : 143 : case BUFFER_LOCK_SHARE_EXCLUSIVE:
5990 : 143 : wait_event = WAIT_EVENT_BUFFER_SHARE_EXCLUSIVE;
5991 : 143 : break;
5992 : 9335 : case BUFFER_LOCK_SHARE:
5993 : 9335 : wait_event = WAIT_EVENT_BUFFER_SHARED;
5994 : 9335 : break;
5995 : : case BUFFER_LOCK_UNLOCK:
5996 : : pg_unreachable();
5997 : :
5998 : : }
5999 : 22230 : pgstat_report_wait_start(wait_event);
6000 : :
6001 : : /*
6002 : : * Wait until awakened.
6003 : : *
6004 : : * It is possible that we get awakened for a reason other than being
6005 : : * signaled by BufferLockWakeup(). If so, loop back and wait again.
6006 : : * Once we've gotten the lock, re-increment the sema by the number of
6007 : : * additional signals received.
6008 : : */
6009 : : for (;;)
6010 : : {
6011 : 22230 : PGSemaphoreLock(MyProc->sem);
6012 [ + - ]: 22230 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
6013 : 22230 : break;
6014 : 0 : extraWaits++;
6015 : : }
6016 : :
6017 : 22230 : pgstat_report_wait_end();
6018 : :
6019 : : /* Retrying, allow BufferLockReleaseSub to release waiters again. */
6020 : 22230 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6021 : : }
6022 : :
6023 : : /* Remember that we now hold this lock */
6024 : 112116182 : entry->data.lockmode = mode;
6025 : :
6026 : : /*
6027 : : * Fix the process wait semaphore's count for any absorbed wakeups.
6028 : : */
6029 [ - + ]: 112116182 : while (unlikely(extraWaits-- > 0))
6030 : 0 : PGSemaphoreUnlock(MyProc->sem);
6031 : 112116182 : }
6032 : :
6033 : : /*
6034 : : * Release a previously acquired buffer content lock.
6035 : : */
6036 : : static void
6037 : 64531462 : BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr)
6038 : : {
6039 : : BufferLockMode mode;
6040 : : uint64 oldstate;
6041 : : uint64 sub;
6042 : :
6043 : 64531462 : mode = BufferLockDisownInternal(buffer, buf_hdr);
6044 : :
6045 : : /*
6046 : : * Release my hold on lock, after that it can immediately be acquired by
6047 : : * others, even if we still have to wakeup other waiters.
6048 : : */
6049 : 64531462 : sub = BufferLockReleaseSub(mode);
6050 : :
6051 : 64531462 : oldstate = pg_atomic_sub_fetch_u64(&buf_hdr->state, sub);
6052 : :
6053 : 64531462 : BufferLockProcessRelease(buf_hdr, mode, oldstate);
6054 : :
6055 : : /*
6056 : : * Now okay to allow cancel/die interrupts.
6057 : : */
6058 : 64531462 : RESUME_INTERRUPTS();
6059 : 64531462 : }
6060 : :
6061 : :
6062 : : /*
6063 : : * Acquire the content lock for the buffer, but only if we don't have to wait.
6064 : : *
6065 : : * It is allowed to try to conditionally acquire a lock on a buffer that this
6066 : : * backend has already locked, but the lock acquisition will always fail, even
6067 : : * if the new lock acquisition does not conflict with an already held lock
6068 : : * (e.g. two share locks). This is because we currently do not have space to
6069 : : * track multiple lock ownerships of the same buffer within one backend. That
6070 : : * is ok for the current uses of BufferLockConditional().
6071 : : */
6072 : : static bool
6073 : 2152417 : BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
6074 : : {
6075 : 2152417 : PrivateRefCountEntry *entry = GetPrivateRefCountEntry(buffer, true);
6076 : : bool mustwait;
6077 : :
6078 : : /*
6079 : : * As described above, if we're trying to lock a buffer this backend
6080 : : * already has locked, return false, independent of the existing and
6081 : : * desired lock level.
6082 : : */
6083 [ - + ]: 2152417 : if (entry->data.lockmode != BUFFER_LOCK_UNLOCK)
6084 : 0 : return false;
6085 : :
6086 : : /*
6087 : : * Lock out cancel/die interrupts until we exit the code section protected
6088 : : * by the content lock. This ensures that interrupts will not interfere
6089 : : * with manipulations of data structures in shared memory.
6090 : : */
6091 : 2152417 : HOLD_INTERRUPTS();
6092 : :
6093 : : /* Check for the lock */
6094 : 2152417 : mustwait = BufferLockAttempt(buf_hdr, mode);
6095 : :
6096 [ + + ]: 2152417 : if (mustwait)
6097 : : {
6098 : : /* Failed to get lock, so release interrupt holdoff */
6099 : 518 : RESUME_INTERRUPTS();
6100 : : }
6101 : : else
6102 : : {
6103 : 2151899 : entry->data.lockmode = mode;
6104 : : }
6105 : :
6106 : 2152417 : return !mustwait;
6107 : : }
6108 : :
6109 : : /*
6110 : : * Internal function that tries to atomically acquire the content lock in the
6111 : : * passed in mode.
6112 : : *
6113 : : * This function will not block waiting for a lock to become free - that's the
6114 : : * caller's job.
6115 : : *
6116 : : * Similar to LWLockAttemptLock().
6117 : : */
6118 : : static inline bool
6119 : 114314697 : BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode)
6120 : : {
6121 : : uint64 old_state;
6122 : :
6123 : : /*
6124 : : * Read once outside the loop, later iterations will get the newer value
6125 : : * via compare & exchange.
6126 : : */
6127 : 114314697 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6128 : :
6129 : : /* loop until we've determined whether we could acquire the lock or not */
6130 : : while (true)
6131 : 16296 : {
6132 : : uint64 desired_state;
6133 : : bool lock_free;
6134 : :
6135 : 114330993 : desired_state = old_state;
6136 : :
6137 [ + + ]: 114330993 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6138 : : {
6139 : 38982791 : lock_free = (old_state & BM_LOCK_MASK) == 0;
6140 [ + + ]: 38982791 : if (lock_free)
6141 : 38955364 : desired_state += BM_LOCK_VAL_EXCLUSIVE;
6142 : : }
6143 [ + + ]: 75348202 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6144 : : {
6145 : 722662 : lock_free = (old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) == 0;
6146 [ + + ]: 722662 : if (lock_free)
6147 : 722375 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
6148 : : }
6149 : : else
6150 : : {
6151 : 74625540 : lock_free = (old_state & BM_LOCK_VAL_EXCLUSIVE) == 0;
6152 [ + + ]: 74625540 : if (lock_free)
6153 : 74606078 : desired_state += BM_LOCK_VAL_SHARED;
6154 : : }
6155 : :
6156 : : /*
6157 : : * Attempt to swap in the state we are expecting. If we didn't see
6158 : : * lock to be free, that's just the old value. If we saw it as free,
6159 : : * we'll attempt to mark it acquired. The reason that we always swap
6160 : : * in the value is that this doubles as a memory barrier. We could try
6161 : : * to be smarter and only swap in values if we saw the lock as free,
6162 : : * but benchmark haven't shown it as beneficial so far.
6163 : : *
6164 : : * Retry if the value changed since we last looked at it.
6165 : : */
6166 [ + + ]: 114330993 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
6167 : : &old_state, desired_state)))
6168 : : {
6169 [ + + ]: 114314697 : if (lock_free)
6170 : : {
6171 : : /* Great! Got the lock. */
6172 : 114268081 : return false;
6173 : : }
6174 : : else
6175 : 46616 : return true; /* somebody else has the lock */
6176 : : }
6177 : : }
6178 : :
6179 : : pg_unreachable();
6180 : : }
6181 : :
6182 : : /*
6183 : : * Add ourselves to the end of the content lock's wait queue.
6184 : : */
6185 : : static void
6186 : 23868 : BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode)
6187 : : {
6188 : : /*
6189 : : * If we don't have a PGPROC structure, there's no way to wait. This
6190 : : * should never occur, since MyProc should only be null during shared
6191 : : * memory initialization.
6192 : : */
6193 [ - + ]: 23868 : if (MyProc == NULL)
6194 [ # # ]: 0 : elog(PANIC, "cannot wait without a PGPROC structure");
6195 : :
6196 [ - + ]: 23868 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
6197 [ # # ]: 0 : elog(PANIC, "queueing for lock while waiting on another one");
6198 : :
6199 : 23868 : LockBufHdr(buf_hdr);
6200 : :
6201 : : /* setting the flag is protected by the spinlock */
6202 : 23868 : pg_atomic_fetch_or_u64(&buf_hdr->state, BM_LOCK_HAS_WAITERS);
6203 : :
6204 : : /*
6205 : : * These are currently used both for lwlocks and buffer content locks,
6206 : : * which is acceptable, although not pretty, because a backend can't wait
6207 : : * for both types of locks at the same time.
6208 : : */
6209 : 23868 : MyProc->lwWaiting = LW_WS_WAITING;
6210 : 23868 : MyProc->lwWaitMode = mode;
6211 : :
6212 : 23868 : proclist_push_tail(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6213 : :
6214 : : /* Can release the mutex now */
6215 : 23868 : UnlockBufHdr(buf_hdr);
6216 : 23868 : }
6217 : :
6218 : : /*
6219 : : * Remove ourselves from the waitlist.
6220 : : *
6221 : : * This is used if we queued ourselves because we thought we needed to sleep
6222 : : * but, after further checking, we discovered that we don't actually need to
6223 : : * do so.
6224 : : */
6225 : : static void
6226 : 1638 : BufferLockDequeueSelf(BufferDesc *buf_hdr)
6227 : : {
6228 : : bool on_waitlist;
6229 : :
6230 : 1638 : LockBufHdr(buf_hdr);
6231 : :
6232 : 1638 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
6233 [ + + ]: 1638 : if (on_waitlist)
6234 : 1134 : proclist_delete(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6235 : :
6236 [ + + ]: 1638 : if (proclist_is_empty(&buf_hdr->lock_waiters) &&
6237 [ + + ]: 1581 : (pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS) != 0)
6238 : : {
6239 : 1078 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_HAS_WAITERS);
6240 : : }
6241 : :
6242 : : /* XXX: combine with fetch_and above? */
6243 : 1638 : UnlockBufHdr(buf_hdr);
6244 : :
6245 : : /* clear waiting state again, nice for debugging */
6246 [ + + ]: 1638 : if (on_waitlist)
6247 : 1134 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
6248 : : else
6249 : : {
6250 : 504 : int extraWaits = 0;
6251 : :
6252 : :
6253 : : /*
6254 : : * Somebody else dequeued us and has or will wake us up. Deal with the
6255 : : * superfluous absorption of a wakeup.
6256 : : */
6257 : :
6258 : : /*
6259 : : * Clear BM_LOCK_WAKE_IN_PROGRESS if somebody woke us before we
6260 : : * removed ourselves - they'll have set it.
6261 : : */
6262 : 504 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6263 : :
6264 : : /*
6265 : : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
6266 : : * get reset at some inconvenient point later. Most of the time this
6267 : : * will immediately return.
6268 : : */
6269 : : for (;;)
6270 : : {
6271 : 504 : PGSemaphoreLock(MyProc->sem);
6272 [ + - ]: 504 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
6273 : 504 : break;
6274 : 0 : extraWaits++;
6275 : : }
6276 : :
6277 : : /*
6278 : : * Fix the process wait semaphore's count for any absorbed wakeups.
6279 : : */
6280 [ - + ]: 504 : while (extraWaits-- > 0)
6281 : 0 : PGSemaphoreUnlock(MyProc->sem);
6282 : : }
6283 : 1638 : }
6284 : :
6285 : : /*
6286 : : * Stop treating lock as held by current backend.
6287 : : *
6288 : : * After calling this function it's the callers responsibility to ensure that
6289 : : * the lock gets released, even in case of an error. This only is desirable if
6290 : : * the lock is going to be released in a different process than the process
6291 : : * that acquired it.
6292 : : */
6293 : : static inline void
6294 : 0 : BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr)
6295 : : {
6296 : 0 : BufferLockDisownInternal(buffer, buf_hdr);
6297 : 0 : RESUME_INTERRUPTS();
6298 : 0 : }
6299 : :
6300 : : /*
6301 : : * Stop treating lock as held by current backend.
6302 : : *
6303 : : * This is the code that can be shared between actually releasing a lock
6304 : : * (BufferLockUnlock()) and just not tracking ownership of the lock anymore
6305 : : * without releasing the lock (BufferLockDisown()).
6306 : : */
6307 : : static inline int
6308 : 114268081 : BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr)
6309 : : {
6310 : : BufferLockMode mode;
6311 : : PrivateRefCountEntry *ref;
6312 : :
6313 : 114268081 : ref = GetPrivateRefCountEntry(buffer, false);
6314 [ - + ]: 114268081 : if (ref == NULL)
6315 [ # # ]: 0 : elog(ERROR, "lock %d is not held", buffer);
6316 : 114268081 : mode = ref->data.lockmode;
6317 : 114268081 : ref->data.lockmode = BUFFER_LOCK_UNLOCK;
6318 : :
6319 : 114268081 : return mode;
6320 : : }
6321 : :
6322 : : /*
6323 : : * Wakeup all the lockers that currently have a chance to acquire the lock.
6324 : : *
6325 : : * wake_exclusive indicates whether exclusive lock waiters should be woken up.
6326 : : */
6327 : : static void
6328 : 22112 : BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive)
6329 : : {
6330 : 22112 : bool new_wake_in_progress = false;
6331 : 22112 : bool wake_share_exclusive = true;
6332 : : proclist_head wakeup;
6333 : : proclist_mutable_iter iter;
6334 : :
6335 : 22112 : proclist_init(&wakeup);
6336 : :
6337 : : /* lock wait list while collecting backends to wake up */
6338 : 22112 : LockBufHdr(buf_hdr);
6339 : :
6340 [ + + + + : 32296 : proclist_foreach_modify(iter, &buf_hdr->lock_waiters, lwWaitLink)
+ + ]
6341 : : {
6342 : 23275 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6343 : :
6344 : : /*
6345 : : * Already woke up a conflicting lock, so skip over this wait list
6346 : : * entry.
6347 : : */
6348 [ + + + + ]: 23275 : if (!wake_exclusive && waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6349 : 544 : continue;
6350 [ + + - + ]: 22731 : if (!wake_share_exclusive && waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6351 : 0 : continue;
6352 : :
6353 : 22731 : proclist_delete(&buf_hdr->lock_waiters, iter.cur, lwWaitLink);
6354 : 22731 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
6355 : :
6356 : : /*
6357 : : * Prevent additional wakeups until retryer gets to run. Backends that
6358 : : * are just waiting for the lock to become free don't retry
6359 : : * automatically.
6360 : : */
6361 : 22731 : new_wake_in_progress = true;
6362 : :
6363 : : /*
6364 : : * Signal that the process isn't on the wait list anymore. This allows
6365 : : * BufferLockDequeueSelf() to remove itself from the waitlist with a
6366 : : * proclist_delete(), rather than having to check if it has been
6367 : : * removed from the list.
6368 : : */
6369 : : Assert(waiter->lwWaiting == LW_WS_WAITING);
6370 : 22731 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
6371 : :
6372 : : /*
6373 : : * Don't wakeup further waiters after waking a conflicting waiter.
6374 : : */
6375 [ + + ]: 22731 : if (waiter->lwWaitMode == BUFFER_LOCK_SHARE)
6376 : : {
6377 : : /*
6378 : : * Share locks conflict with exclusive locks.
6379 : : */
6380 : 9494 : wake_exclusive = false;
6381 : : }
6382 [ + + ]: 13237 : else if (waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6383 : : {
6384 : : /*
6385 : : * Share-exclusive locks conflict with share-exclusive and
6386 : : * exclusive locks.
6387 : : */
6388 : 146 : wake_exclusive = false;
6389 : 146 : wake_share_exclusive = false;
6390 : : }
6391 [ + - ]: 13091 : else if (waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6392 : : {
6393 : : /*
6394 : : * Exclusive locks conflict with all other locks, there's no point
6395 : : * in waking up anybody else.
6396 : : */
6397 : 13091 : break;
6398 : : }
6399 : : }
6400 : :
6401 : : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS);
6402 : :
6403 : : /* unset required flags, and release lock, in one fell swoop */
6404 : : {
6405 : : uint64 old_state;
6406 : : uint64 desired_state;
6407 : :
6408 : 22112 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6409 : : while (true)
6410 : : {
6411 : 22146 : desired_state = old_state;
6412 : :
6413 : : /* compute desired flags */
6414 : :
6415 [ + + ]: 22146 : if (new_wake_in_progress)
6416 : 21847 : desired_state |= BM_LOCK_WAKE_IN_PROGRESS;
6417 : : else
6418 : 299 : desired_state &= ~BM_LOCK_WAKE_IN_PROGRESS;
6419 : :
6420 [ + + ]: 22146 : if (proclist_is_empty(&buf_hdr->lock_waiters))
6421 : 20249 : desired_state &= ~BM_LOCK_HAS_WAITERS;
6422 : :
6423 : 22146 : desired_state &= ~BM_LOCKED; /* release lock */
6424 : :
6425 [ + + ]: 22146 : if (pg_atomic_compare_exchange_u64(&buf_hdr->state, &old_state,
6426 : : desired_state))
6427 : 22112 : break;
6428 : : }
6429 : : }
6430 : :
6431 : : /* Awaken any waiters I removed from the queue. */
6432 [ + + + + : 44843 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
+ + ]
6433 : : {
6434 : 22731 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6435 : :
6436 : 22731 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
6437 : :
6438 : : /*
6439 : : * Guarantee that lwWaiting being unset only becomes visible once the
6440 : : * unlink from the link has completed. Otherwise the target backend
6441 : : * could be woken up for other reason and enqueue for a new lock - if
6442 : : * that happens before the list unlink happens, the list would end up
6443 : : * being corrupted.
6444 : : *
6445 : : * The barrier pairs with the LockBufHdr() when enqueuing for another
6446 : : * lock.
6447 : : */
6448 : 22731 : pg_write_barrier();
6449 : 22731 : waiter->lwWaiting = LW_WS_NOT_WAITING;
6450 : 22731 : PGSemaphoreUnlock(waiter->sem);
6451 : : }
6452 : 22112 : }
6453 : :
6454 : : /*
6455 : : * Compute subtraction from buffer state for a release of a held lock in
6456 : : * `mode`.
6457 : : *
6458 : : * This is separated from BufferLockUnlock() as we want to combine the lock
6459 : : * release with other atomic operations when possible, leading to the lock
6460 : : * release being done in multiple places, each needing to compute what to
6461 : : * subtract from the lock state.
6462 : : */
6463 : : static inline uint64
6464 : 114268081 : BufferLockReleaseSub(BufferLockMode mode)
6465 : : {
6466 : : /*
6467 : : * Turns out that a switch() leads gcc to generate sufficiently worse code
6468 : : * for this to show up in profiles...
6469 : : */
6470 [ + + ]: 114268081 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6471 : 38954843 : return BM_LOCK_VAL_EXCLUSIVE;
6472 [ + + ]: 75313238 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6473 : 4130825 : return BM_LOCK_VAL_SHARE_EXCLUSIVE;
6474 : : else
6475 : : {
6476 : : Assert(mode == BUFFER_LOCK_SHARE);
6477 : 71182413 : return BM_LOCK_VAL_SHARED;
6478 : : }
6479 : :
6480 : : return 0; /* keep compiler quiet */
6481 : : }
6482 : :
6483 : : /*
6484 : : * Handle work that needs to be done after releasing a lock that was held in
6485 : : * `mode`, where `lockstate` is the result of the atomic operation modifying
6486 : : * the state variable.
6487 : : *
6488 : : * This is separated from BufferLockUnlock() as we want to combine the lock
6489 : : * release with other atomic operations when possible, leading to the lock
6490 : : * release being done in multiple places.
6491 : : */
6492 : : static void
6493 : 114268081 : BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate)
6494 : : {
6495 : 114268081 : bool check_waiters = false;
6496 : 114268081 : bool wake_exclusive = false;
6497 : :
6498 : : /* nobody else can have that kind of lock */
6499 : : Assert(!(lockstate & BM_LOCK_VAL_EXCLUSIVE));
6500 : :
6501 : : /*
6502 : : * If we're still waiting for backends to get scheduled, don't wake them
6503 : : * up again. Otherwise check if we need to look through the waitqueue to
6504 : : * wake other backends.
6505 : : */
6506 [ + + ]: 114268081 : if ((lockstate & BM_LOCK_HAS_WAITERS) &&
6507 [ + + ]: 80148 : !(lockstate & BM_LOCK_WAKE_IN_PROGRESS))
6508 : : {
6509 [ + + ]: 41930 : if ((lockstate & BM_LOCK_MASK) == 0)
6510 : : {
6511 : : /*
6512 : : * We released a lock and the lock was, in that moment, free. We
6513 : : * therefore can wake waiters for any kind of lock.
6514 : : */
6515 : 22112 : check_waiters = true;
6516 : 22112 : wake_exclusive = true;
6517 : : }
6518 [ - + ]: 19818 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6519 : : {
6520 : : /*
6521 : : * We released the lock, but another backend still holds a lock.
6522 : : * We can't have released an exclusive lock, as there couldn't
6523 : : * have been other lock holders. If we released a share lock, no
6524 : : * waiters need to be woken up, as there must be other share
6525 : : * lockers. However, if we held a share-exclusive lock, another
6526 : : * backend now could acquire a share-exclusive lock.
6527 : : */
6528 : 0 : check_waiters = true;
6529 : 0 : wake_exclusive = false;
6530 : : }
6531 : : }
6532 : :
6533 : : /*
6534 : : * As waking up waiters requires the spinlock to be acquired, only do so
6535 : : * if necessary.
6536 : : */
6537 [ + + ]: 114268081 : if (check_waiters)
6538 : 22112 : BufferLockWakeup(buf_hdr, wake_exclusive);
6539 : 114268081 : }
6540 : :
6541 : : /*
6542 : : * BufferLockHeldByMeInMode - test whether my process holds the content lock
6543 : : * in the specified mode
6544 : : *
6545 : : * This is meant as debug support only.
6546 : : */
6547 : : static bool
6548 : 0 : BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode)
6549 : : {
6550 : : PrivateRefCountEntry *entry =
6551 : 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6552 : :
6553 [ # # ]: 0 : if (!entry)
6554 : 0 : return false;
6555 : : else
6556 : 0 : return entry->data.lockmode == mode;
6557 : : }
6558 : :
6559 : : /*
6560 : : * BufferLockHeldByMe - test whether my process holds the content lock in any
6561 : : * mode
6562 : : *
6563 : : * This is meant as debug support only.
6564 : : */
6565 : : static bool
6566 : 0 : BufferLockHeldByMe(BufferDesc *buf_hdr)
6567 : : {
6568 : : PrivateRefCountEntry *entry =
6569 : 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6570 : :
6571 [ # # ]: 0 : if (!entry)
6572 : 0 : return false;
6573 : : else
6574 : 0 : return entry->data.lockmode != BUFFER_LOCK_UNLOCK;
6575 : : }
6576 : :
6577 : : /*
6578 : : * Release the content lock for the buffer.
6579 : : */
6580 : : void
6581 : 69387636 : UnlockBuffer(Buffer buffer)
6582 : : {
6583 : : BufferDesc *buf_hdr;
6584 : :
6585 : : Assert(BufferIsPinned(buffer));
6586 [ + + ]: 69387636 : if (BufferIsLocal(buffer))
6587 : 5200643 : return; /* local buffers need no lock */
6588 : :
6589 : 64186993 : buf_hdr = GetBufferDescriptor(buffer - 1);
6590 : 64186993 : BufferLockUnlock(buffer, buf_hdr);
6591 : : }
6592 : :
6593 : : /*
6594 : : * Acquire the content_lock for the buffer.
6595 : : */
6596 : : void
6597 : 118242664 : LockBufferInternal(Buffer buffer, BufferLockMode mode)
6598 : : {
6599 : : BufferDesc *buf_hdr;
6600 : :
6601 : : /*
6602 : : * We can't wait if we haven't got a PGPROC. This should only occur
6603 : : * during bootstrap or shared memory initialization. Put an Assert here
6604 : : * to catch unsafe coding practices.
6605 : : */
6606 : : Assert(!(MyProc == NULL && IsUnderPostmaster));
6607 : :
6608 : : /* handled in LockBuffer() wrapper */
6609 : : Assert(mode != BUFFER_LOCK_UNLOCK);
6610 : :
6611 : : Assert(BufferIsPinned(buffer));
6612 [ + + ]: 118242664 : if (BufferIsLocal(buffer))
6613 : 6470837 : return; /* local buffers need no lock */
6614 : :
6615 : 111771827 : buf_hdr = GetBufferDescriptor(buffer - 1);
6616 : :
6617 : : /*
6618 : : * Test the most frequent lock modes first. While a switch (mode) would be
6619 : : * nice, at least gcc generates considerably worse code for it.
6620 : : *
6621 : : * Call BufferLockAcquire() with a constant argument for mode, to generate
6622 : : * more efficient code for the different lock modes.
6623 : : */
6624 [ + + ]: 111771827 : if (mode == BUFFER_LOCK_SHARE)
6625 : 74590868 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE);
6626 [ + - ]: 37180959 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
6627 : 37180959 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_EXCLUSIVE);
6628 [ # # ]: 0 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6629 : 0 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE);
6630 : : else
6631 [ # # ]: 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
6632 : : }
6633 : :
6634 : : /*
6635 : : * Acquire the content_lock for the buffer, but only if we don't have to wait.
6636 : : *
6637 : : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
6638 : : */
6639 : : bool
6640 : 1860654 : ConditionalLockBuffer(Buffer buffer)
6641 : : {
6642 : : BufferDesc *buf;
6643 : :
6644 : : Assert(BufferIsPinned(buffer));
6645 [ + + ]: 1860654 : if (BufferIsLocal(buffer))
6646 : 86268 : return true; /* act as though we got it */
6647 : :
6648 : 1774386 : buf = GetBufferDescriptor(buffer - 1);
6649 : :
6650 : 1774386 : return BufferLockConditional(buffer, buf, BUFFER_LOCK_EXCLUSIVE);
6651 : : }
6652 : :
6653 : : /*
6654 : : * Verify that this backend is pinning the buffer exactly once.
6655 : : *
6656 : : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
6657 : : * holds a pin on the buffer. We do not care whether some other backend does.
6658 : : */
6659 : : void
6660 : 2703195 : CheckBufferIsPinnedOnce(Buffer buffer)
6661 : : {
6662 [ + + ]: 2703195 : if (BufferIsLocal(buffer))
6663 : : {
6664 [ - + ]: 1049 : if (LocalRefCount[-buffer - 1] != 1)
6665 [ # # ]: 0 : elog(ERROR, "incorrect local pin count: %d",
6666 : : LocalRefCount[-buffer - 1]);
6667 : : }
6668 : : else
6669 : : {
6670 [ - + ]: 2702146 : if (GetPrivateRefCount(buffer) != 1)
6671 [ # # ]: 0 : elog(ERROR, "incorrect local pin count: %d",
6672 : : GetPrivateRefCount(buffer));
6673 : : }
6674 : 2703195 : }
6675 : :
6676 : : /*
6677 : : * LockBufferForCleanup - lock a buffer in preparation for deleting items
6678 : : *
6679 : : * Items may be deleted from a disk page only when the caller (a) holds an
6680 : : * exclusive lock on the buffer and (b) has observed that no other backend
6681 : : * holds a pin on the buffer. If there is a pin, then the other backend
6682 : : * might have a pointer into the buffer (for example, a heapscan reference
6683 : : * to an item --- see README for more details). It's OK if a pin is added
6684 : : * after the cleanup starts, however; the newly-arrived backend will be
6685 : : * unable to look at the page until we release the exclusive lock.
6686 : : *
6687 : : * To implement this protocol, a would-be deleter must pin the buffer and
6688 : : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
6689 : : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
6690 : : * it has successfully observed pin count = 1.
6691 : : */
6692 : : void
6693 : 28560 : LockBufferForCleanup(Buffer buffer)
6694 : : {
6695 : : BufferDesc *bufHdr;
6696 : 28560 : TimestampTz waitStart = 0;
6697 : 28560 : bool waiting = false;
6698 : 28560 : bool logged_recovery_conflict = false;
6699 : :
6700 : : Assert(BufferIsPinned(buffer));
6701 : : Assert(PinCountWaitBuf == NULL);
6702 : :
6703 : 28560 : CheckBufferIsPinnedOnce(buffer);
6704 : :
6705 : : /*
6706 : : * We do not yet need to be worried about in-progress AIOs holding a pin,
6707 : : * as we, so far, only support doing reads via AIO and this function can
6708 : : * only be called once the buffer is valid (i.e. no read can be in
6709 : : * flight).
6710 : : */
6711 : :
6712 : : /* Nobody else to wait for */
6713 [ + + ]: 28560 : if (BufferIsLocal(buffer))
6714 : 18 : return;
6715 : :
6716 : 28542 : bufHdr = GetBufferDescriptor(buffer - 1);
6717 : :
6718 : : for (;;)
6719 : 11 : {
6720 : : uint64 buf_state;
6721 : 28553 : uint64 unset_bits = 0;
6722 : :
6723 : : /* Try to acquire lock */
6724 : 28553 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6725 : 28553 : buf_state = LockBufHdr(bufHdr);
6726 : :
6727 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6728 [ + + ]: 28553 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6729 : : {
6730 : : /* Successfully acquired exclusive lock with pincount 1 */
6731 : 28542 : UnlockBufHdr(bufHdr);
6732 : 28542 : goto cleanup_lock_acquired;
6733 : : }
6734 : : /* Failed, so mark myself as waiting for pincount 1 */
6735 [ - + ]: 11 : if (buf_state & BM_PIN_COUNT_WAITER)
6736 : : {
6737 : 0 : UnlockBufHdr(bufHdr);
6738 : 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6739 [ # # ]: 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
6740 : : }
6741 : 11 : bufHdr->wait_backend_pgprocno = MyProcNumber;
6742 : 11 : PinCountWaitBuf = bufHdr;
6743 : :
6744 : : /*
6745 : : * Publish BM_PIN_COUNT_WAITER while retaining the buffer header lock.
6746 : : * The shared refcount can be decremented while BM_LOCKED is set, so
6747 : : * use an atomic operation that preserves concurrent refcount changes.
6748 : : */
6749 : 11 : pg_atomic_fetch_or_u64(&bufHdr->state, BM_PIN_COUNT_WAITER);
6750 : :
6751 : : /*
6752 : : * Recheck the refcount after publishing the waiter flag, while shared
6753 : : * refcount increments are still prevented by BM_LOCKED. If only our
6754 : : * pin remains, the cleanup-lock condition has already been satisfied,
6755 : : * so remove the waiter state and return without sleeping.
6756 : : */
6757 : 11 : buf_state = pg_atomic_read_u64(&bufHdr->state);
6758 : :
6759 [ - + ]: 11 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6760 : : {
6761 : 0 : UnlockBufHdrExt(bufHdr, buf_state,
6762 : : 0, BM_PIN_COUNT_WAITER,
6763 : : 0);
6764 : 0 : PinCountWaitBuf = NULL;
6765 : 0 : goto cleanup_lock_acquired;
6766 : : }
6767 : :
6768 : 11 : UnlockBufHdr(bufHdr);
6769 : 11 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6770 : :
6771 : : /* Wait to be signaled by UnpinBuffer() */
6772 [ + + ]: 11 : if (InHotStandby)
6773 : : {
6774 [ + + ]: 8 : if (!waiting)
6775 : : {
6776 : : /* adjust the process title to indicate that it's waiting */
6777 : 2 : set_ps_display_suffix("waiting");
6778 : 2 : waiting = true;
6779 : : }
6780 : :
6781 : : /*
6782 : : * Emit the log message if the startup process is waiting longer
6783 : : * than deadlock_timeout for recovery conflict on buffer pin.
6784 : : *
6785 : : * Skip this if first time through because the startup process has
6786 : : * not started waiting yet in this case. So, the wait start
6787 : : * timestamp is set after this logic.
6788 : : */
6789 [ + + + + ]: 8 : if (waitStart != 0 && !logged_recovery_conflict)
6790 : : {
6791 : 3 : TimestampTz now = GetCurrentTimestamp();
6792 : :
6793 [ + + ]: 3 : if (TimestampDifferenceExceeds(waitStart, now,
6794 : : DeadlockTimeout))
6795 : : {
6796 : 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6797 : : waitStart, now, NULL, true);
6798 : 2 : logged_recovery_conflict = true;
6799 : : }
6800 : : }
6801 : :
6802 : : /*
6803 : : * Set the wait start timestamp if logging is enabled and first
6804 : : * time through.
6805 : : */
6806 [ + - + + ]: 8 : if (log_recovery_conflict_waits && waitStart == 0)
6807 : 2 : waitStart = GetCurrentTimestamp();
6808 : :
6809 : : /* Publish the bufid that Startup process waits on */
6810 : 8 : SetStartupBufferPinWaitBufId(buffer - 1);
6811 : : /* Set alarm and then wait to be signaled by UnpinBuffer() */
6812 : 8 : ResolveRecoveryConflictWithBufferPin();
6813 : : /* Reset the published bufid */
6814 : 8 : SetStartupBufferPinWaitBufId(-1);
6815 : : }
6816 : : else
6817 : 3 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
6818 : :
6819 : : /*
6820 : : * Remove flag marking us as waiter. Normally this will not be set
6821 : : * anymore, but ProcWaitForSignal() can return for other signals as
6822 : : * well. We take care to only reset the flag if we're the waiter, as
6823 : : * theoretically another backend could have started waiting. That's
6824 : : * impossible with the current usages due to table level locking, but
6825 : : * better be safe.
6826 : : */
6827 : 11 : buf_state = LockBufHdr(bufHdr);
6828 [ + + ]: 11 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
6829 [ + - ]: 6 : bufHdr->wait_backend_pgprocno == MyProcNumber)
6830 : 6 : unset_bits |= BM_PIN_COUNT_WAITER;
6831 : :
6832 : 11 : UnlockBufHdrExt(bufHdr, buf_state,
6833 : : 0, unset_bits,
6834 : : 0);
6835 : :
6836 : 11 : PinCountWaitBuf = NULL;
6837 : : /* Loop back and try again */
6838 : : }
6839 : :
6840 : 28542 : cleanup_lock_acquired:
6841 : :
6842 : : /*
6843 : : * Emit the log message if recovery conflict on buffer pin was resolved
6844 : : * but the startup process waited longer than deadlock_timeout for it.
6845 : : */
6846 [ + + ]: 28542 : if (logged_recovery_conflict)
6847 : 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6848 : : waitStart, GetCurrentTimestamp(),
6849 : : NULL, false);
6850 : :
6851 [ + + ]: 28542 : if (waiting)
6852 : : {
6853 : : /* reset ps display to remove the suffix if we added one */
6854 : 2 : set_ps_display_remove_suffix();
6855 : 2 : waiting = false;
6856 : : }
6857 : :
6858 : 28542 : return;
6859 : : }
6860 : :
6861 : : /*
6862 : : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
6863 : : * requests cancellation of all pin holders that are blocking it.
6864 : : */
6865 : : bool
6866 : 3 : HoldingBufferPinThatDelaysRecovery(void)
6867 : : {
6868 : 3 : int bufid = GetStartupBufferPinWaitBufId();
6869 : :
6870 : : /*
6871 : : * If we get woken slowly then it's possible that the Startup process was
6872 : : * already woken by other backends before we got here. Also possible that
6873 : : * we get here by multiple interrupts or interrupts at inappropriate
6874 : : * times, so make sure we do nothing if the bufid is not set.
6875 : : */
6876 [ + + ]: 3 : if (bufid < 0)
6877 : 1 : return false;
6878 : :
6879 [ + - ]: 2 : if (GetPrivateRefCount(bufid + 1) > 0)
6880 : 2 : return true;
6881 : :
6882 : 0 : return false;
6883 : : }
6884 : :
6885 : : /*
6886 : : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
6887 : : *
6888 : : * We won't loop, but just check once to see if the pin count is OK. If
6889 : : * not, return false with no lock held.
6890 : : */
6891 : : bool
6892 : 619930 : ConditionalLockBufferForCleanup(Buffer buffer)
6893 : : {
6894 : : BufferDesc *bufHdr;
6895 : : uint64 buf_state,
6896 : : refcount;
6897 : :
6898 : : Assert(BufferIsValid(buffer));
6899 : :
6900 : : /* see AIO related comment in LockBufferForCleanup() */
6901 : :
6902 [ + + ]: 619930 : if (BufferIsLocal(buffer))
6903 : : {
6904 : 12373 : refcount = LocalRefCount[-buffer - 1];
6905 : : /* There should be exactly one pin */
6906 : : Assert(refcount > 0);
6907 [ + + ]: 12373 : if (refcount != 1)
6908 : 1540 : return false;
6909 : : /* Nobody else to wait for */
6910 : 10833 : return true;
6911 : : }
6912 : :
6913 : : /* There should be exactly one local pin */
6914 : 607557 : refcount = GetPrivateRefCount(buffer);
6915 : : Assert(refcount);
6916 [ + + ]: 607557 : if (refcount != 1)
6917 : 495 : return false;
6918 : :
6919 : : /* Try to acquire lock */
6920 [ + + ]: 607062 : if (!ConditionalLockBuffer(buffer))
6921 : 71 : return false;
6922 : :
6923 : 606991 : bufHdr = GetBufferDescriptor(buffer - 1);
6924 : 606991 : buf_state = LockBufHdr(bufHdr);
6925 : 606991 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
6926 : :
6927 : : Assert(refcount > 0);
6928 [ + + ]: 606991 : if (refcount == 1)
6929 : : {
6930 : : /* Successfully acquired exclusive lock with pincount 1 */
6931 : 606835 : UnlockBufHdr(bufHdr);
6932 : 606835 : return true;
6933 : : }
6934 : :
6935 : : /* Failed, so release the lock */
6936 : 156 : UnlockBufHdr(bufHdr);
6937 : 156 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6938 : 156 : return false;
6939 : : }
6940 : :
6941 : : /*
6942 : : * IsBufferCleanupOK - as above, but we already have the lock
6943 : : *
6944 : : * Check whether it's OK to perform cleanup on a buffer we've already
6945 : : * locked. If we observe that the pin count is 1, our exclusive lock
6946 : : * happens to be a cleanup lock, and we can proceed with anything that
6947 : : * would have been allowable had we sought a cleanup lock originally.
6948 : : */
6949 : : bool
6950 : 2708 : IsBufferCleanupOK(Buffer buffer)
6951 : : {
6952 : : BufferDesc *bufHdr;
6953 : : uint64 buf_state;
6954 : :
6955 : : Assert(BufferIsValid(buffer));
6956 : :
6957 : : /* see AIO related comment in LockBufferForCleanup() */
6958 : :
6959 [ - + ]: 2708 : if (BufferIsLocal(buffer))
6960 : : {
6961 : : /* There should be exactly one pin */
6962 [ # # ]: 0 : if (LocalRefCount[-buffer - 1] != 1)
6963 : 0 : return false;
6964 : : /* Nobody else to wait for */
6965 : 0 : return true;
6966 : : }
6967 : :
6968 : : /* There should be exactly one local pin */
6969 [ - + ]: 2708 : if (GetPrivateRefCount(buffer) != 1)
6970 : 0 : return false;
6971 : :
6972 : 2708 : bufHdr = GetBufferDescriptor(buffer - 1);
6973 : :
6974 : : /* caller must hold exclusive lock on buffer */
6975 : : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
6976 : :
6977 : 2708 : buf_state = LockBufHdr(bufHdr);
6978 : :
6979 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6980 [ + - ]: 2708 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6981 : : {
6982 : : /* pincount is OK. */
6983 : 2708 : UnlockBufHdr(bufHdr);
6984 : 2708 : return true;
6985 : : }
6986 : :
6987 : 0 : UnlockBufHdr(bufHdr);
6988 : 0 : return false;
6989 : : }
6990 : :
6991 : : /*
6992 : : * Helper for BufferBeginSetHintBits() and BufferSetHintBits16().
6993 : : *
6994 : : * This checks if the current lock mode already suffices to allow hint bits
6995 : : * being set and, if not, whether the current lock can be upgraded.
6996 : : *
6997 : : * Updates *lockstate when returning true.
6998 : : */
6999 : : static inline bool
7000 : 15196141 : SharedBufferBeginSetHintBits(Buffer buffer, BufferDesc *buf_hdr, uint64 *lockstate)
7001 : : {
7002 : : uint64 old_state;
7003 : : PrivateRefCountEntry *ref;
7004 : : BufferLockMode mode;
7005 : :
7006 : 15196141 : ref = GetPrivateRefCountEntry(buffer, true);
7007 : :
7008 [ - + ]: 15196141 : if (ref == NULL)
7009 [ # # ]: 0 : elog(ERROR, "buffer is not pinned");
7010 : :
7011 : 15196141 : mode = ref->data.lockmode;
7012 [ - + ]: 15196141 : if (mode == BUFFER_LOCK_UNLOCK)
7013 [ # # ]: 0 : elog(ERROR, "buffer is not locked");
7014 : :
7015 : : /* we're done if we are already holding a sufficient lock level */
7016 [ + + + + ]: 15196141 : if (mode == BUFFER_LOCK_EXCLUSIVE || mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
7017 : : {
7018 : 11787517 : *lockstate = pg_atomic_read_u64(&buf_hdr->state);
7019 : 11787517 : return true;
7020 : : }
7021 : :
7022 : : /*
7023 : : * We are only holding a share lock right now, try to upgrade it to
7024 : : * SHARE_EXCLUSIVE.
7025 : : */
7026 : : Assert(mode == BUFFER_LOCK_SHARE);
7027 : :
7028 : 3408624 : old_state = pg_atomic_read_u64(&buf_hdr->state);
7029 : : while (true)
7030 : 12 : {
7031 : : uint64 desired_state;
7032 : :
7033 : 3408636 : desired_state = old_state;
7034 : :
7035 : : /*
7036 : : * Can't upgrade if somebody else holds the lock in exclusive or
7037 : : * share-exclusive mode.
7038 : : */
7039 [ + + ]: 3408636 : if (unlikely((old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) != 0))
7040 : : {
7041 : 169 : return false;
7042 : : }
7043 : :
7044 : : /* currently held lock state */
7045 : 3408467 : desired_state -= BM_LOCK_VAL_SHARED;
7046 : :
7047 : : /* new lock level */
7048 : 3408467 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
7049 : :
7050 [ + + ]: 3408467 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
7051 : : &old_state, desired_state)))
7052 : : {
7053 : 3408455 : ref->data.lockmode = BUFFER_LOCK_SHARE_EXCLUSIVE;
7054 : 3408455 : *lockstate = desired_state;
7055 : :
7056 : 3408455 : return true;
7057 : : }
7058 : : }
7059 : : }
7060 : :
7061 : : /*
7062 : : * Try to acquire the right to set hint bits on the buffer.
7063 : : *
7064 : : * To be allowed to set hint bits, this backend needs to hold either a
7065 : : * share-exclusive or an exclusive lock. In case this backend only holds a
7066 : : * share lock, this function will try to upgrade the lock to
7067 : : * share-exclusive. The caller is only allowed to set hint bits if true is
7068 : : * returned.
7069 : : *
7070 : : * Once BufferBeginSetHintBits() has returned true, hint bits may be set
7071 : : * without further calls to BufferBeginSetHintBits(), until the buffer is
7072 : : * unlocked.
7073 : : *
7074 : : *
7075 : : * Requiring a share-exclusive lock to set hint bits prevents setting hint
7076 : : * bits on buffers that are currently being written out, which could corrupt
7077 : : * the checksum on the page. Flushing buffers also requires a share-exclusive
7078 : : * lock.
7079 : : *
7080 : : * Due to a lock >= share-exclusive being required to set hint bits, only one
7081 : : * backend can set hint bits at a time. Allowing multiple backends to set hint
7082 : : * bits would require more complicated locking: For setting hint bits we'd
7083 : : * need to store the count of backends currently setting hint bits, for I/O we
7084 : : * would need another lock-level conflicting with the hint-setting
7085 : : * lock-level. Given that the share-exclusive lock for setting hint bits is
7086 : : * only held for a short time, that backends often would just set the same
7087 : : * hint bits and that the cost of occasionally not setting hint bits in hotly
7088 : : * accessed pages is fairly low, this seems like an acceptable tradeoff.
7089 : : */
7090 : : bool
7091 : 426169 : BufferBeginSetHintBits(Buffer buffer)
7092 : : {
7093 : : BufferDesc *buf_hdr;
7094 : : uint64 lockstate;
7095 : :
7096 [ + + ]: 426169 : if (BufferIsLocal(buffer))
7097 : : {
7098 : : /*
7099 : : * NB: Will need to check if there is a write in progress, once it is
7100 : : * possible for writes to be done asynchronously.
7101 : : */
7102 : 2405 : return true;
7103 : : }
7104 : :
7105 : 423764 : buf_hdr = GetBufferDescriptor(buffer - 1);
7106 : :
7107 : 423764 : return SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate);
7108 : : }
7109 : :
7110 : : /*
7111 : : * End a phase of setting hint bits on this buffer, started with
7112 : : * BufferBeginSetHintBits().
7113 : : *
7114 : : * This would strictly speaking not be required (i.e. the caller could do
7115 : : * MarkBufferDirtyHint() if so desired), but allows us to perform some sanity
7116 : : * checks.
7117 : : */
7118 : : void
7119 : 426159 : BufferFinishSetHintBits(Buffer buffer, bool mark_dirty, bool buffer_std)
7120 : : {
7121 : : if (!BufferIsLocal(buffer))
7122 : : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
7123 : : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
7124 : :
7125 [ + + ]: 426159 : if (mark_dirty)
7126 : 234307 : MarkBufferDirtyHint(buffer, buffer_std);
7127 : 426159 : }
7128 : :
7129 : : /*
7130 : : * Try to set hint bits on a single 16bit value in a buffer.
7131 : : *
7132 : : * If hint bits are allowed to be set, set *ptr = val, try to mark the buffer
7133 : : * dirty and return true. Otherwise false is returned.
7134 : : *
7135 : : * *ptr needs to be a pointer to memory within the buffer.
7136 : : *
7137 : : * This is a bit faster than BufferBeginSetHintBits() /
7138 : : * BufferFinishSetHintBits() when setting hints once in a buffer, but slower
7139 : : * than the former when setting hint bits multiple times in the same buffer.
7140 : : */
7141 : : bool
7142 : 15576038 : BufferSetHintBits16(uint16 *ptr, uint16 val, Buffer buffer)
7143 : : {
7144 : : BufferDesc *buf_hdr;
7145 : : uint64 lockstate;
7146 : : #ifdef USE_ASSERT_CHECKING
7147 : : char *page;
7148 : :
7149 : : /* verify that the address is on the page */
7150 : : page = BufferGetPage(buffer);
7151 : : Assert((char *) ptr >= page && (char *) ptr < (page + BLCKSZ));
7152 : : #endif
7153 : :
7154 [ + + ]: 15576038 : if (BufferIsLocal(buffer))
7155 : : {
7156 : 803661 : *ptr = val;
7157 : :
7158 : 803661 : MarkLocalBufferDirty(buffer);
7159 : :
7160 : 803661 : return true;
7161 : : }
7162 : :
7163 : 14772377 : buf_hdr = GetBufferDescriptor(buffer - 1);
7164 : :
7165 [ + + ]: 14772377 : if (SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate))
7166 : : {
7167 : 14772218 : *ptr = val;
7168 : :
7169 : 14772218 : MarkSharedBufferDirtyHint(buffer, buf_hdr, lockstate, true);
7170 : :
7171 : 14772218 : return true;
7172 : : }
7173 : :
7174 : 159 : return false;
7175 : : }
7176 : :
7177 : :
7178 : : /*
7179 : : * Functions for buffer I/O handling
7180 : : *
7181 : : * Also note that these are used only for shared buffers, not local ones.
7182 : : */
7183 : :
7184 : : /*
7185 : : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
7186 : : */
7187 : : static void
7188 : 182 : WaitIO(BufferDesc *buf)
7189 : : {
7190 : 182 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
7191 : :
7192 : : /*
7193 : : * Should never end up here with unsubmitted IO, as no AIO unaware code
7194 : : * may be used while in batch mode and AIO aware code needs to have
7195 : : * submitted all staged IO to avoid deadlocks & slowness.
7196 : : */
7197 : : Assert(!pgaio_have_staged());
7198 : :
7199 : 182 : ConditionVariablePrepareToSleep(cv);
7200 : : for (;;)
7201 : 182 : {
7202 : : uint64 buf_state;
7203 : : PgAioWaitRef iow;
7204 : :
7205 : : /*
7206 : : * It may not be necessary to acquire the spinlock to check the flag
7207 : : * here, but since this test is essential for correctness, we'd better
7208 : : * play it safe.
7209 : : */
7210 : 364 : buf_state = LockBufHdr(buf);
7211 : :
7212 : : /*
7213 : : * Copy the wait reference while holding the spinlock. This protects
7214 : : * against a concurrent TerminateBufferIO() in another backend from
7215 : : * clearing the wref while it's being read.
7216 : : */
7217 : 364 : iow = buf->io_wref;
7218 : 364 : UnlockBufHdr(buf);
7219 : :
7220 : : /* no IO in progress, we don't need to wait */
7221 [ + + ]: 364 : if (!(buf_state & BM_IO_IN_PROGRESS))
7222 : 182 : break;
7223 : :
7224 : : /*
7225 : : * The buffer has asynchronous IO in progress, wait for it to
7226 : : * complete.
7227 : : */
7228 [ + + ]: 182 : if (pgaio_wref_valid(&iow))
7229 : : {
7230 : 25 : pgaio_wref_wait(&iow);
7231 : :
7232 : : /*
7233 : : * The AIO subsystem internally uses condition variables and thus
7234 : : * might remove this backend from the BufferDesc's CV. While that
7235 : : * wouldn't cause a correctness issue (the first CV sleep just
7236 : : * immediately returns if not already registered), it seems worth
7237 : : * avoiding unnecessary loop iterations, given that we take care
7238 : : * to do so at the start of the function.
7239 : : */
7240 : 25 : ConditionVariablePrepareToSleep(cv);
7241 : 25 : continue;
7242 : : }
7243 : :
7244 : : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
7245 : 157 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
7246 : : }
7247 : 182 : ConditionVariableCancelSleep();
7248 : 182 : }
7249 : :
7250 : : /*
7251 : : * StartSharedBufferIO: begin I/O on this buffer
7252 : : * (Assumptions)
7253 : : * The buffer is Pinned
7254 : : *
7255 : : * In several scenarios the buffer may already be undergoing I/O in this or
7256 : : * another backend. How to best handle that depends on the caller's
7257 : : * situation. It might be appropriate to wait synchronously (e.g., because the
7258 : : * buffer is about to be invalidated); wait asynchronously, using the buffer's
7259 : : * IO wait reference (e.g., because the caller is doing readahead and doesn't
7260 : : * need the buffer to be ready immediately); or to not wait at all (e.g.,
7261 : : * because the caller is trying to combine IO for this buffer with another
7262 : : * buffer).
7263 : : *
7264 : : * How and whether to wait is controlled by the wait and io_wref
7265 : : * parameters. In detail:
7266 : : *
7267 : : * - If the caller passes a non-NULL io_wref and the buffer has an I/O wait
7268 : : * reference, the *io_wref is set to the buffer's io_wref and
7269 : : * BUFFER_IO_IN_PROGRESS is returned. This is done regardless of the wait
7270 : : * parameter.
7271 : : *
7272 : : * - If the caller passes a NULL io_wref (i.e. the caller does not want to
7273 : : * asynchronously wait for the completion of the IO), wait = false and the
7274 : : * buffer is undergoing IO, BUFFER_IO_IN_PROGRESS is returned.
7275 : : *
7276 : : * - If wait = true and either the buffer does not have a wait reference,
7277 : : * or the caller passes io_wref = NULL, WaitIO() is used to wait for the IO
7278 : : * to complete. To avoid the potential of deadlocks and unnecessary delays,
7279 : : * all staged I/O is submitted before waiting.
7280 : : *
7281 : : * Input operations are only attempted on buffers that are not BM_VALID, and
7282 : : * output operations only on buffers that are BM_VALID and BM_DIRTY, so we can
7283 : : * always tell if the work is already done. If no I/O is necessary,
7284 : : * BUFFER_IO_ALREADY_DONE is returned.
7285 : : *
7286 : : * If we successfully marked the buffer as BM_IO_IN_PROGRESS,
7287 : : * BUFFER_IO_READY_FOR_IO is returned.
7288 : : */
7289 : : StartBufferIOResult
7290 : 2873260 : StartSharedBufferIO(BufferDesc *buf, bool forInput, bool wait, PgAioWaitRef *io_wref)
7291 : : {
7292 : : uint64 buf_state;
7293 : :
7294 : 2873260 : ResourceOwnerEnlarge(CurrentResourceOwner);
7295 : :
7296 : : for (;;)
7297 : : {
7298 : 2873442 : buf_state = LockBufHdr(buf);
7299 : :
7300 [ + + ]: 2873442 : if (!(buf_state & BM_IO_IN_PROGRESS))
7301 : 2871242 : break;
7302 : :
7303 : : /* Join the existing IO */
7304 [ + + + + ]: 2200 : if (io_wref != NULL && pgaio_wref_valid(&buf->io_wref))
7305 : : {
7306 : 2010 : *io_wref = buf->io_wref;
7307 : 2010 : UnlockBufHdr(buf);
7308 : :
7309 : 2010 : return BUFFER_IO_IN_PROGRESS;
7310 : : }
7311 [ + + ]: 190 : else if (!wait)
7312 : : {
7313 : 8 : UnlockBufHdr(buf);
7314 : 8 : return BUFFER_IO_IN_PROGRESS;
7315 : : }
7316 : : else
7317 : : {
7318 : : /*
7319 : : * With wait = true, we always have to wait if the caller has
7320 : : * passed io_wref = NULL.
7321 : : *
7322 : : * Even with io_wref != NULL, we have to wait if the buffer's wait
7323 : : * ref is not valid but the IO is in progress, someone else
7324 : : * started IO but hasn't set the wait ref yet. We have no choice
7325 : : * but to wait until the IO completes.
7326 : : */
7327 : 182 : UnlockBufHdr(buf);
7328 : :
7329 : : /*
7330 : : * If this backend currently has staged IO, submit it before
7331 : : * waiting for in-progress IO, to avoid potential deadlocks and
7332 : : * unnecessary delays.
7333 : : */
7334 : 182 : pgaio_submit_staged();
7335 : :
7336 : 182 : WaitIO(buf);
7337 : : }
7338 : : }
7339 : :
7340 : : /* Once we get here, there is definitely no I/O active on this buffer */
7341 : :
7342 : : /* Check if someone else already did the I/O */
7343 [ + + + + ]: 2871242 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
7344 : : {
7345 : 338 : UnlockBufHdr(buf);
7346 : 338 : return BUFFER_IO_ALREADY_DONE;
7347 : : }
7348 : :
7349 : : /*
7350 : : * No IO in progress and not already done; we will start IO. It's possible
7351 : : * that the IO was in progress but we're not done, because the IO errored
7352 : : * out. We'll do the IO ourselves.
7353 : : */
7354 : 2870904 : UnlockBufHdrExt(buf, buf_state,
7355 : : BM_IO_IN_PROGRESS, 0,
7356 : : 0);
7357 : :
7358 : 2870904 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
7359 : : BufferDescriptorGetBuffer(buf));
7360 : :
7361 : 2870904 : return BUFFER_IO_READY_FOR_IO;
7362 : : }
7363 : :
7364 : : /*
7365 : : * Wrapper around StartSharedBufferIO / StartLocalBufferIO. Only to be used
7366 : : * when the caller doesn't otherwise need to care about local vs shared. See
7367 : : * StartSharedBufferIO() for details.
7368 : : */
7369 : : StartBufferIOResult
7370 : 1584660 : StartBufferIO(Buffer buffer, bool forInput, bool wait, PgAioWaitRef *io_wref)
7371 : : {
7372 : : BufferDesc *buf_hdr;
7373 : :
7374 [ + + ]: 1584660 : if (BufferIsLocal(buffer))
7375 : : {
7376 : 11017 : buf_hdr = GetLocalBufferDescriptor(-buffer - 1);
7377 : :
7378 : 11017 : return StartLocalBufferIO(buf_hdr, forInput, wait, io_wref);
7379 : : }
7380 : : else
7381 : : {
7382 : 1573643 : buf_hdr = GetBufferDescriptor(buffer - 1);
7383 : :
7384 : 1573643 : return StartSharedBufferIO(buf_hdr, forInput, wait, io_wref);
7385 : : }
7386 : : }
7387 : :
7388 : : /*
7389 : : * TerminateBufferIO: release a buffer we were doing I/O on
7390 : : * (Assumptions)
7391 : : * My process is executing IO for the buffer
7392 : : * BM_IO_IN_PROGRESS bit is set for the buffer
7393 : : * The buffer is Pinned
7394 : : *
7395 : : * If clear_dirty is true, we clear the buffer's BM_DIRTY flag. This is
7396 : : * appropriate when terminating a successful write.
7397 : : *
7398 : : * set_flag_bits gets ORed into the buffer's flags. It must include
7399 : : * BM_IO_ERROR in a failure case. For successful completion it could
7400 : : * be 0, or BM_VALID if we just finished reading in the page.
7401 : : *
7402 : : * If forget_owner is true, we release the buffer I/O from the current
7403 : : * resource owner. (forget_owner=false is used when the resource owner itself
7404 : : * is being released)
7405 : : */
7406 : : void
7407 : 2696719 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint64 set_flag_bits,
7408 : : bool forget_owner, bool release_aio)
7409 : : {
7410 : : uint64 buf_state;
7411 : 2696719 : uint64 unset_flag_bits = 0;
7412 : 2696719 : int refcount_change = 0;
7413 : :
7414 : 2696719 : buf_state = LockBufHdr(buf);
7415 : :
7416 : : Assert(buf_state & BM_IO_IN_PROGRESS);
7417 : 2696719 : unset_flag_bits |= BM_IO_IN_PROGRESS;
7418 : :
7419 : : /* Clear earlier errors, if this IO failed, it'll be marked again */
7420 : 2696719 : unset_flag_bits |= BM_IO_ERROR;
7421 : :
7422 [ + + ]: 2696719 : if (clear_dirty)
7423 : 700401 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
7424 : :
7425 [ + + ]: 2696719 : if (release_aio)
7426 : : {
7427 : : /* release ownership by the AIO subsystem */
7428 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
7429 : 1397207 : refcount_change = -1;
7430 : 1397207 : pgaio_wref_clear(&buf->io_wref);
7431 : : }
7432 : :
7433 : 2696719 : buf_state = UnlockBufHdrExt(buf, buf_state,
7434 : : set_flag_bits, unset_flag_bits,
7435 : : refcount_change);
7436 : :
7437 [ + + ]: 2696719 : if (forget_owner)
7438 : 1299489 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
7439 : : BufferDescriptorGetBuffer(buf));
7440 : :
7441 : 2696719 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
7442 : :
7443 : : /*
7444 : : * Support LockBufferForCleanup()
7445 : : *
7446 : : * We may have just released the last pin other than the waiter's. In most
7447 : : * cases, this backend holds another pin on the buffer. But, if, for
7448 : : * example, this backend is completing an IO issued by another backend, it
7449 : : * may be time to wake the waiter.
7450 : : */
7451 [ + + - + ]: 2696719 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
7452 : 0 : WakePinCountWaiter(buf);
7453 : 2696719 : }
7454 : :
7455 : : /*
7456 : : * AbortBufferIO: Clean up active buffer I/O after an error.
7457 : : *
7458 : : * All LWLocks & content locks we might have held have been released, but we
7459 : : * haven't yet released buffer pins, so the buffer is still pinned.
7460 : : *
7461 : : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
7462 : : * possible the error condition wasn't related to the I/O.
7463 : : *
7464 : : * Note: this does not remove the buffer I/O from the resource owner.
7465 : : * That's correct when we're releasing the whole resource owner, but
7466 : : * beware if you use this in other contexts.
7467 : : */
7468 : : static void
7469 : 15 : AbortBufferIO(Buffer buffer)
7470 : : {
7471 : 15 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
7472 : : uint64 buf_state;
7473 : :
7474 : 15 : buf_state = LockBufHdr(buf_hdr);
7475 : : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
7476 : :
7477 [ + - ]: 15 : if (!(buf_state & BM_VALID))
7478 : : {
7479 : : Assert(!(buf_state & BM_DIRTY));
7480 : 15 : UnlockBufHdr(buf_hdr);
7481 : : }
7482 : : else
7483 : : {
7484 : : Assert(buf_state & BM_DIRTY);
7485 : 0 : UnlockBufHdr(buf_hdr);
7486 : :
7487 : : /* Issue notice if this is not the first failure... */
7488 [ # # ]: 0 : if (buf_state & BM_IO_ERROR)
7489 : : {
7490 : : /* Buffer is pinned, so we can read tag without spinlock */
7491 [ # # ]: 0 : ereport(WARNING,
7492 : : (errcode(ERRCODE_IO_ERROR),
7493 : : errmsg("could not write block %u of %s",
7494 : : buf_hdr->tag.blockNum,
7495 : : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
7496 : : BufTagGetForkNum(&buf_hdr->tag)).str),
7497 : : errdetail("Multiple failures --- write error might be permanent.")));
7498 : : }
7499 : : }
7500 : :
7501 : 15 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
7502 : 15 : }
7503 : :
7504 : : /*
7505 : : * Error context callback for errors occurring during shared buffer writes.
7506 : : */
7507 : : static void
7508 : 44 : shared_buffer_write_error_callback(void *arg)
7509 : : {
7510 : 44 : BufferDesc *bufHdr = (BufferDesc *) arg;
7511 : :
7512 : : /* Buffer is pinned, so we can read the tag without locking the spinlock */
7513 [ + - ]: 44 : if (bufHdr != NULL)
7514 : 88 : errcontext("writing block %u of relation \"%s\"",
7515 : : bufHdr->tag.blockNum,
7516 : 44 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
7517 : : BufTagGetForkNum(&bufHdr->tag)).str);
7518 : 44 : }
7519 : :
7520 : : /*
7521 : : * Error context callback for errors occurring during local buffer writes.
7522 : : */
7523 : : static void
7524 : 0 : local_buffer_write_error_callback(void *arg)
7525 : : {
7526 : 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
7527 : :
7528 [ # # ]: 0 : if (bufHdr != NULL)
7529 : 0 : errcontext("writing block %u of relation \"%s\"",
7530 : : bufHdr->tag.blockNum,
7531 : 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
7532 : : MyProcNumber,
7533 : : BufTagGetForkNum(&bufHdr->tag)).str);
7534 : 0 : }
7535 : :
7536 : : /*
7537 : : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
7538 : : */
7539 : : static int
7540 : 13255628 : rlocator_comparator(const void *p1, const void *p2)
7541 : : {
7542 : 13255628 : RelFileLocator n1 = *(const RelFileLocator *) p1;
7543 : 13255628 : RelFileLocator n2 = *(const RelFileLocator *) p2;
7544 : :
7545 [ + + ]: 13255628 : if (n1.relNumber < n2.relNumber)
7546 : 13212024 : return -1;
7547 [ + + ]: 43604 : else if (n1.relNumber > n2.relNumber)
7548 : 41786 : return 1;
7549 : :
7550 [ - + ]: 1818 : if (n1.dbOid < n2.dbOid)
7551 : 0 : return -1;
7552 [ - + ]: 1818 : else if (n1.dbOid > n2.dbOid)
7553 : 0 : return 1;
7554 : :
7555 [ - + ]: 1818 : if (n1.spcOid < n2.spcOid)
7556 : 0 : return -1;
7557 [ - + ]: 1818 : else if (n1.spcOid > n2.spcOid)
7558 : 0 : return 1;
7559 : : else
7560 : 1818 : return 0;
7561 : : }
7562 : :
7563 : : /*
7564 : : * Lock buffer header - set BM_LOCKED in buffer state.
7565 : : */
7566 : : uint64
7567 : 28198130 : LockBufHdr(BufferDesc *desc)
7568 : : {
7569 : : uint64 old_buf_state;
7570 : :
7571 : : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
7572 : :
7573 : : while (true)
7574 : : {
7575 : : /*
7576 : : * Always try once to acquire the lock directly, without setting up
7577 : : * the spin-delay infrastructure. The work necessary for that shows up
7578 : : * in profiles and is rarely necessary.
7579 : : */
7580 : 28199035 : old_buf_state = pg_atomic_fetch_or_u64(&desc->state, BM_LOCKED);
7581 [ + + ]: 28199035 : if (likely(!(old_buf_state & BM_LOCKED)))
7582 : 28198130 : break; /* got lock */
7583 : :
7584 : : /* and then spin without atomic operations until lock is released */
7585 : : {
7586 : : SpinDelayStatus delayStatus;
7587 : :
7588 : 905 : init_local_spin_delay(&delayStatus);
7589 : :
7590 [ + + ]: 3974 : while (old_buf_state & BM_LOCKED)
7591 : : {
7592 : 3069 : perform_spin_delay(&delayStatus);
7593 : 3069 : old_buf_state = pg_atomic_read_u64(&desc->state);
7594 : : }
7595 : 905 : finish_spin_delay(&delayStatus);
7596 : : }
7597 : :
7598 : : /*
7599 : : * Retry. The lock might obviously already be re-acquired by the time
7600 : : * we're attempting to get it again.
7601 : : */
7602 : : }
7603 : :
7604 : 28198130 : return old_buf_state | BM_LOCKED;
7605 : : }
7606 : :
7607 : : /*
7608 : : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
7609 : : * state at that point.
7610 : : *
7611 : : * Obviously the buffer could be locked by the time the value is returned, so
7612 : : * this is primarily useful in CAS style loops.
7613 : : */
7614 : : pg_noinline uint64
7615 : 648 : WaitBufHdrUnlocked(BufferDesc *buf)
7616 : : {
7617 : : SpinDelayStatus delayStatus;
7618 : : uint64 buf_state;
7619 : :
7620 : 648 : init_local_spin_delay(&delayStatus);
7621 : :
7622 : 648 : buf_state = pg_atomic_read_u64(&buf->state);
7623 : :
7624 [ + + ]: 878 : while (buf_state & BM_LOCKED)
7625 : : {
7626 : 230 : perform_spin_delay(&delayStatus);
7627 : 230 : buf_state = pg_atomic_read_u64(&buf->state);
7628 : : }
7629 : :
7630 : 648 : finish_spin_delay(&delayStatus);
7631 : :
7632 : 648 : return buf_state;
7633 : : }
7634 : :
7635 : : /*
7636 : : * BufferTag comparator.
7637 : : */
7638 : : static inline int
7639 : 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
7640 : : {
7641 : : int ret;
7642 : : RelFileLocator rlocatora;
7643 : : RelFileLocator rlocatorb;
7644 : :
7645 : 0 : rlocatora = BufTagGetRelFileLocator(ba);
7646 : 0 : rlocatorb = BufTagGetRelFileLocator(bb);
7647 : :
7648 : 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
7649 : :
7650 [ # # ]: 0 : if (ret != 0)
7651 : 0 : return ret;
7652 : :
7653 [ # # ]: 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
7654 : 0 : return -1;
7655 [ # # ]: 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
7656 : 0 : return 1;
7657 : :
7658 [ # # ]: 0 : if (ba->blockNum < bb->blockNum)
7659 : 0 : return -1;
7660 [ # # ]: 0 : if (ba->blockNum > bb->blockNum)
7661 : 0 : return 1;
7662 : :
7663 : 0 : return 0;
7664 : : }
7665 : :
7666 : : /*
7667 : : * Comparator determining the writeout order in a checkpoint.
7668 : : *
7669 : : * It is important that tablespaces are compared first, the logic balancing
7670 : : * writes between tablespaces relies on it.
7671 : : */
7672 : : static inline int
7673 : 3416812 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
7674 : : {
7675 : : /* compare tablespace */
7676 [ + + ]: 3416812 : if (a->tsId < b->tsId)
7677 : 9476 : return -1;
7678 [ + + ]: 3407336 : else if (a->tsId > b->tsId)
7679 : 30084 : return 1;
7680 : : /* compare relation */
7681 [ + + ]: 3377252 : if (a->relNumber < b->relNumber)
7682 : 965396 : return -1;
7683 [ + + ]: 2411856 : else if (a->relNumber > b->relNumber)
7684 : 908471 : return 1;
7685 : : /* compare fork */
7686 [ + + ]: 1503385 : else if (a->forkNum < b->forkNum)
7687 : 69526 : return -1;
7688 [ + + ]: 1433859 : else if (a->forkNum > b->forkNum)
7689 : 69988 : return 1;
7690 : : /* compare block number */
7691 [ + + ]: 1363871 : else if (a->blockNum < b->blockNum)
7692 : 664123 : return -1;
7693 [ + + ]: 699748 : else if (a->blockNum > b->blockNum)
7694 : 653248 : return 1;
7695 : : /* equal page IDs are unlikely, but not impossible */
7696 : 46500 : return 0;
7697 : : }
7698 : :
7699 : : /*
7700 : : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
7701 : : * progress.
7702 : : */
7703 : : static int
7704 : 281609 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
7705 : : {
7706 : 281609 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
7707 : 281609 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
7708 : :
7709 : : /* we want a min-heap, so return 1 for the a < b */
7710 [ + + ]: 281609 : if (sa->progress < sb->progress)
7711 : 257909 : return 1;
7712 [ + + ]: 23700 : else if (sa->progress == sb->progress)
7713 : 1056 : return 0;
7714 : : else
7715 : 22644 : return -1;
7716 : : }
7717 : :
7718 : : /*
7719 : : * Initialize a writeback context, discarding potential previous state.
7720 : : *
7721 : : * *max_pending is a pointer instead of an immediate value, so the coalesce
7722 : : * limits can easily changed by the GUC mechanism, and so calling code does
7723 : : * not have to check the current configuration. A value of 0 means that no
7724 : : * writeback control will be performed.
7725 : : */
7726 : : void
7727 : 3108 : WritebackContextInit(WritebackContext *context, int *max_pending)
7728 : : {
7729 : : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7730 : :
7731 : 3108 : context->max_pending = max_pending;
7732 : 3108 : context->nr_pending = 0;
7733 : 3108 : }
7734 : :
7735 : : /*
7736 : : * Add buffer to list of pending writeback requests.
7737 : : */
7738 : : void
7739 : 696596 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
7740 : : BufferTag *tag)
7741 : : {
7742 : : PendingWriteback *pending;
7743 : :
7744 : : /*
7745 : : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
7746 : : * point in tracking in that case.
7747 : : */
7748 [ + + ]: 696596 : if (io_direct_flags & IO_DIRECT_DATA ||
7749 [ + + ]: 696077 : !enableFsync)
7750 : 696594 : return;
7751 : :
7752 : : /*
7753 : : * Add buffer to the pending writeback array, unless writeback control is
7754 : : * disabled.
7755 : : */
7756 [ - + ]: 2 : if (*wb_context->max_pending > 0)
7757 : : {
7758 : : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7759 : :
7760 : 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
7761 : :
7762 : 0 : pending->tag = *tag;
7763 : : }
7764 : :
7765 : : /*
7766 : : * Perform pending flushes if the writeback limit is exceeded. This
7767 : : * includes the case where previously an item has been added, but control
7768 : : * is now disabled.
7769 : : */
7770 [ + - ]: 2 : if (wb_context->nr_pending >= *wb_context->max_pending)
7771 : 2 : IssuePendingWritebacks(wb_context, io_context);
7772 : : }
7773 : :
7774 : : #define ST_SORT sort_pending_writebacks
7775 : : #define ST_ELEMENT_TYPE PendingWriteback
7776 : : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
7777 : : #define ST_SCOPE static
7778 : : #define ST_DEFINE
7779 : : #include "lib/sort_template.h"
7780 : :
7781 : : /*
7782 : : * Issue all pending writeback requests, previously scheduled with
7783 : : * ScheduleBufferTagForWriteback, to the OS.
7784 : : *
7785 : : * Because this is only used to improve the OSs IO scheduling we try to never
7786 : : * error out - it's just a hint.
7787 : : */
7788 : : void
7789 : 1219 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
7790 : : {
7791 : : instr_time io_start;
7792 : : int i;
7793 : :
7794 [ + - ]: 1219 : if (wb_context->nr_pending == 0)
7795 : 1219 : return;
7796 : :
7797 : : /*
7798 : : * Executing the writes in-order can make them a lot faster, and allows to
7799 : : * merge writeback requests to consecutive blocks into larger writebacks.
7800 : : */
7801 : 0 : sort_pending_writebacks(wb_context->pending_writebacks,
7802 : 0 : wb_context->nr_pending);
7803 : :
7804 : 0 : io_start = pgstat_prepare_io_time(track_io_timing);
7805 : :
7806 : : /*
7807 : : * Coalesce neighbouring writes, but nothing else. For that we iterate
7808 : : * through the, now sorted, array of pending flushes, and look forward to
7809 : : * find all neighbouring (or identical) writes.
7810 : : */
7811 [ # # ]: 0 : for (i = 0; i < wb_context->nr_pending; i++)
7812 : : {
7813 : : PendingWriteback *cur;
7814 : : PendingWriteback *next;
7815 : : SMgrRelation reln;
7816 : : int ahead;
7817 : : BufferTag tag;
7818 : : RelFileLocator currlocator;
7819 : 0 : Size nblocks = 1;
7820 : :
7821 : 0 : cur = &wb_context->pending_writebacks[i];
7822 : 0 : tag = cur->tag;
7823 : 0 : currlocator = BufTagGetRelFileLocator(&tag);
7824 : :
7825 : : /*
7826 : : * Peek ahead, into following writeback requests, to see if they can
7827 : : * be combined with the current one.
7828 : : */
7829 [ # # ]: 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
7830 : : {
7831 : :
7832 : 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
7833 : :
7834 : : /* different file, stop */
7835 [ # # # # : 0 : if (!RelFileLocatorEquals(currlocator,
# # ]
7836 [ # # ]: 0 : BufTagGetRelFileLocator(&next->tag)) ||
7837 : 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
7838 : : break;
7839 : :
7840 : : /* ok, block queued twice, skip */
7841 [ # # ]: 0 : if (cur->tag.blockNum == next->tag.blockNum)
7842 : 0 : continue;
7843 : :
7844 : : /* only merge consecutive writes */
7845 [ # # ]: 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
7846 : 0 : break;
7847 : :
7848 : 0 : nblocks++;
7849 : 0 : cur = next;
7850 : : }
7851 : :
7852 : 0 : i += ahead;
7853 : :
7854 : : /* and finally tell the kernel to write the data to storage */
7855 : 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
7856 : 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
7857 : : }
7858 : :
7859 : : /*
7860 : : * Assume that writeback requests are only issued for buffers containing
7861 : : * blocks of permanent relations.
7862 : : */
7863 : 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
7864 : 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
7865 : :
7866 : 0 : wb_context->nr_pending = 0;
7867 : : }
7868 : :
7869 : : /* ResourceOwner callbacks */
7870 : :
7871 : : static void
7872 : 15 : ResOwnerReleaseBufferIO(Datum res)
7873 : : {
7874 : 15 : Buffer buffer = DatumGetInt32(res);
7875 : :
7876 : 15 : AbortBufferIO(buffer);
7877 : 15 : }
7878 : :
7879 : : static char *
7880 : 0 : ResOwnerPrintBufferIO(Datum res)
7881 : : {
7882 : 0 : Buffer buffer = DatumGetInt32(res);
7883 : :
7884 : 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
7885 : : }
7886 : :
7887 : : /*
7888 : : * Release buffer as part of resource owner cleanup. This will only be called
7889 : : * if the buffer is pinned. If this backend held the content lock at the time
7890 : : * of the error we also need to release that (note that it is not possible to
7891 : : * hold a content lock without a pin).
7892 : : */
7893 : : static void
7894 : 10732 : ResOwnerReleaseBuffer(Datum res)
7895 : : {
7896 : 10732 : Buffer buffer = DatumGetInt32(res);
7897 : :
7898 : : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
7899 [ - + ]: 10732 : if (!BufferIsValid(buffer))
7900 [ # # ]: 0 : elog(ERROR, "bad buffer ID: %d", buffer);
7901 : :
7902 [ + + ]: 10732 : if (BufferIsLocal(buffer))
7903 : 3995 : UnpinLocalBufferNoOwner(buffer);
7904 : : else
7905 : : {
7906 : : PrivateRefCountEntry *ref;
7907 : :
7908 : 6737 : ref = GetPrivateRefCountEntry(buffer, false);
7909 : :
7910 : : /* not having a private refcount would imply resowner corruption */
7911 : : Assert(ref != NULL);
7912 : :
7913 : : /*
7914 : : * If the buffer was locked at the time of the resowner release,
7915 : : * release the lock now. This should only happen after errors.
7916 : : */
7917 [ + + ]: 6737 : if (ref->data.lockmode != BUFFER_LOCK_UNLOCK)
7918 : : {
7919 : 114 : BufferDesc *buf = GetBufferDescriptor(buffer - 1);
7920 : :
7921 : 114 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
7922 : 114 : BufferLockUnlock(buffer, buf);
7923 : : }
7924 : :
7925 : 6737 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
7926 : : }
7927 : 10732 : }
7928 : :
7929 : : static char *
7930 : 0 : ResOwnerPrintBuffer(Datum res)
7931 : : {
7932 : 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
7933 : : }
7934 : :
7935 : : /*
7936 : : * Helper function to evict unpinned buffer whose buffer header lock is
7937 : : * already acquired.
7938 : : */
7939 : : static bool
7940 : 2518 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
7941 : : {
7942 : : uint64 buf_state;
7943 : : bool result;
7944 : :
7945 : 2518 : *buffer_flushed = false;
7946 : :
7947 : 2518 : buf_state = pg_atomic_read_u64(&(desc->state));
7948 : : Assert(buf_state & BM_LOCKED);
7949 : :
7950 [ - + ]: 2518 : if ((buf_state & BM_VALID) == 0)
7951 : : {
7952 : 0 : UnlockBufHdr(desc);
7953 : 0 : return false;
7954 : : }
7955 : :
7956 : : /* Check that it's not pinned already. */
7957 [ - + ]: 2518 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
7958 : : {
7959 : 0 : UnlockBufHdr(desc);
7960 : 0 : return false;
7961 : : }
7962 : :
7963 : 2518 : PinBuffer_Locked(desc); /* releases spinlock */
7964 : :
7965 : : /* If it was dirty, try to clean it once. */
7966 [ + + ]: 2518 : if (buf_state & BM_DIRTY)
7967 : : {
7968 : 1077 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
7969 : 1077 : *buffer_flushed = true;
7970 : : }
7971 : :
7972 : : /* This will return false if it becomes dirty or someone else pins it. */
7973 : 2518 : result = InvalidateVictimBuffer(desc);
7974 : :
7975 : 2518 : UnpinBuffer(desc);
7976 : :
7977 : 2518 : return result;
7978 : : }
7979 : :
7980 : : /*
7981 : : * Try to evict the current block in a shared buffer.
7982 : : *
7983 : : * This function is intended for testing/development use only!
7984 : : *
7985 : : * To succeed, the buffer must not be pinned on entry, so if the caller had a
7986 : : * particular block in mind, it might already have been replaced by some other
7987 : : * block by the time this function runs. It's also unpinned on return, so the
7988 : : * buffer might be occupied again by the time control is returned, potentially
7989 : : * even by the same block. This inherent raciness without other interlocking
7990 : : * makes the function unsuitable for non-testing usage.
7991 : : *
7992 : : * *buffer_flushed is set to true if the buffer was dirty and has been
7993 : : * flushed, false otherwise. However, *buffer_flushed=true does not
7994 : : * necessarily mean that we flushed the buffer, it could have been flushed by
7995 : : * someone else.
7996 : : *
7997 : : * Returns true if the buffer was valid and it has now been made invalid.
7998 : : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
7999 : : * or if the buffer becomes dirty again while we're trying to write it out.
8000 : : */
8001 : : bool
8002 : 137 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
8003 : : {
8004 : : BufferDesc *desc;
8005 : :
8006 : : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
8007 : :
8008 : : /* Make sure we can pin the buffer. */
8009 : 137 : ResourceOwnerEnlarge(CurrentResourceOwner);
8010 : 137 : ReservePrivateRefCountEntry();
8011 : :
8012 : 137 : desc = GetBufferDescriptor(buf - 1);
8013 : 137 : LockBufHdr(desc);
8014 : :
8015 : 137 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
8016 : : }
8017 : :
8018 : : /*
8019 : : * Try to evict all the shared buffers.
8020 : : *
8021 : : * This function is intended for testing/development use only! See
8022 : : * EvictUnpinnedBuffer().
8023 : : *
8024 : : * The buffers_* parameters are mandatory and indicate the total count of
8025 : : * buffers that:
8026 : : * - buffers_evicted - were evicted
8027 : : * - buffers_flushed - were flushed
8028 : : * - buffers_skipped - could not be evicted
8029 : : */
8030 : : void
8031 : 1 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
8032 : : int32 *buffers_skipped)
8033 : : {
8034 : 1 : *buffers_evicted = 0;
8035 : 1 : *buffers_skipped = 0;
8036 : 1 : *buffers_flushed = 0;
8037 : :
8038 [ + + ]: 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8039 : : {
8040 : 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8041 : : uint64 buf_state;
8042 : : bool buffer_flushed;
8043 : :
8044 [ - + ]: 16384 : CHECK_FOR_INTERRUPTS();
8045 : :
8046 : 16384 : buf_state = pg_atomic_read_u64(&desc->state);
8047 [ + + ]: 16384 : if (!(buf_state & BM_VALID))
8048 : 14351 : continue;
8049 : :
8050 : 2033 : ResourceOwnerEnlarge(CurrentResourceOwner);
8051 : 2033 : ReservePrivateRefCountEntry();
8052 : :
8053 : 2033 : LockBufHdr(desc);
8054 : :
8055 [ + - ]: 2033 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8056 : 2033 : (*buffers_evicted)++;
8057 : : else
8058 : 0 : (*buffers_skipped)++;
8059 : :
8060 [ + + ]: 2033 : if (buffer_flushed)
8061 : 980 : (*buffers_flushed)++;
8062 : : }
8063 : 1 : }
8064 : :
8065 : : /*
8066 : : * Try to evict all the shared buffers containing provided relation's pages.
8067 : : *
8068 : : * This function is intended for testing/development use only! See
8069 : : * EvictUnpinnedBuffer().
8070 : : *
8071 : : * The caller must hold at least AccessShareLock on the relation to prevent
8072 : : * the relation from being dropped.
8073 : : *
8074 : : * The buffers_* parameters are mandatory and indicate the total count of
8075 : : * buffers that:
8076 : : * - buffers_evicted - were evicted
8077 : : * - buffers_flushed - were flushed
8078 : : * - buffers_skipped - could not be evicted
8079 : : */
8080 : : void
8081 : 29 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
8082 : : int32 *buffers_flushed, int32 *buffers_skipped)
8083 : : {
8084 : : Assert(!RelationUsesLocalBuffers(rel));
8085 : :
8086 : 29 : *buffers_skipped = 0;
8087 : 29 : *buffers_evicted = 0;
8088 : 29 : *buffers_flushed = 0;
8089 : :
8090 [ + + ]: 475165 : for (int buf = 1; buf <= NBuffers; buf++)
8091 : : {
8092 : 475136 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8093 : 475136 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8094 : : bool buffer_flushed;
8095 : :
8096 [ - + ]: 475136 : CHECK_FOR_INTERRUPTS();
8097 : :
8098 : : /* An unlocked precheck should be safe and saves some cycles. */
8099 [ + + ]: 475136 : if ((buf_state & BM_VALID) == 0 ||
8100 [ + + ]: 63071 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8101 : 474788 : continue;
8102 : :
8103 : : /* Make sure we can pin the buffer. */
8104 : 348 : ResourceOwnerEnlarge(CurrentResourceOwner);
8105 : 348 : ReservePrivateRefCountEntry();
8106 : :
8107 : 348 : buf_state = LockBufHdr(desc);
8108 : :
8109 : : /* recheck, could have changed without the lock */
8110 [ + - ]: 348 : if ((buf_state & BM_VALID) == 0 ||
8111 [ - + ]: 348 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8112 : : {
8113 : 0 : UnlockBufHdr(desc);
8114 : 0 : continue;
8115 : : }
8116 : :
8117 [ + - ]: 348 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8118 : 348 : (*buffers_evicted)++;
8119 : : else
8120 : 0 : (*buffers_skipped)++;
8121 : :
8122 [ + + ]: 348 : if (buffer_flushed)
8123 : 78 : (*buffers_flushed)++;
8124 : : }
8125 : 29 : }
8126 : :
8127 : : /*
8128 : : * Helper function to mark unpinned buffer dirty whose buffer header lock is
8129 : : * already acquired.
8130 : : */
8131 : : static bool
8132 : 36 : MarkDirtyUnpinnedBufferInternal(Buffer buf, BufferDesc *desc,
8133 : : bool *buffer_already_dirty)
8134 : : {
8135 : : uint64 buf_state;
8136 : 36 : bool result = false;
8137 : :
8138 : 36 : *buffer_already_dirty = false;
8139 : :
8140 : 36 : buf_state = pg_atomic_read_u64(&(desc->state));
8141 : : Assert(buf_state & BM_LOCKED);
8142 : :
8143 [ + + ]: 36 : if ((buf_state & BM_VALID) == 0)
8144 : : {
8145 : 1 : UnlockBufHdr(desc);
8146 : 1 : return false;
8147 : : }
8148 : :
8149 : : /* Check that it's not pinned already. */
8150 [ - + ]: 35 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
8151 : : {
8152 : 0 : UnlockBufHdr(desc);
8153 : 0 : return false;
8154 : : }
8155 : :
8156 : : /* Pin the buffer and then release the buffer spinlock */
8157 : 35 : PinBuffer_Locked(desc);
8158 : :
8159 : : /* If it was not already dirty, mark it as dirty. */
8160 [ + + ]: 35 : if (!(buf_state & BM_DIRTY))
8161 : : {
8162 : 16 : BufferLockAcquire(buf, desc, BUFFER_LOCK_EXCLUSIVE);
8163 : 16 : MarkBufferDirty(buf);
8164 : 16 : result = true;
8165 : 16 : BufferLockUnlock(buf, desc);
8166 : : }
8167 : : else
8168 : 19 : *buffer_already_dirty = true;
8169 : :
8170 : 35 : UnpinBuffer(desc);
8171 : :
8172 : 35 : return result;
8173 : : }
8174 : :
8175 : : /*
8176 : : * Try to mark the provided shared buffer as dirty.
8177 : : *
8178 : : * This function is intended for testing/development use only!
8179 : : *
8180 : : * Same as EvictUnpinnedBuffer() but with MarkBufferDirty() call inside.
8181 : : *
8182 : : * The buffer_already_dirty parameter is mandatory and indicate if the buffer
8183 : : * could not be dirtied because it is already dirty.
8184 : : *
8185 : : * Returns true if the buffer has successfully been marked as dirty.
8186 : : */
8187 : : bool
8188 : 1 : MarkDirtyUnpinnedBuffer(Buffer buf, bool *buffer_already_dirty)
8189 : : {
8190 : : BufferDesc *desc;
8191 : 1 : bool buffer_dirtied = false;
8192 : :
8193 : : Assert(!BufferIsLocal(buf));
8194 : :
8195 : : /* Make sure we can pin the buffer. */
8196 : 1 : ResourceOwnerEnlarge(CurrentResourceOwner);
8197 : 1 : ReservePrivateRefCountEntry();
8198 : :
8199 : 1 : desc = GetBufferDescriptor(buf - 1);
8200 : 1 : LockBufHdr(desc);
8201 : :
8202 : 1 : buffer_dirtied = MarkDirtyUnpinnedBufferInternal(buf, desc, buffer_already_dirty);
8203 : : /* Both can not be true at the same time */
8204 : : Assert(!(buffer_dirtied && *buffer_already_dirty));
8205 : :
8206 : 1 : return buffer_dirtied;
8207 : : }
8208 : :
8209 : : /*
8210 : : * Try to mark all the shared buffers containing provided relation's pages as
8211 : : * dirty.
8212 : : *
8213 : : * This function is intended for testing/development use only! See
8214 : : * MarkDirtyUnpinnedBuffer().
8215 : : *
8216 : : * The buffers_* parameters are mandatory and indicate the total count of
8217 : : * buffers that:
8218 : : * - buffers_dirtied - were dirtied
8219 : : * - buffers_already_dirty - were already dirty
8220 : : * - buffers_skipped - could not be dirtied because of a reason different
8221 : : * than a buffer being already dirty.
8222 : : */
8223 : : void
8224 : 1 : MarkDirtyRelUnpinnedBuffers(Relation rel,
8225 : : int32 *buffers_dirtied,
8226 : : int32 *buffers_already_dirty,
8227 : : int32 *buffers_skipped)
8228 : : {
8229 : : Assert(!RelationUsesLocalBuffers(rel));
8230 : :
8231 : 1 : *buffers_dirtied = 0;
8232 : 1 : *buffers_already_dirty = 0;
8233 : 1 : *buffers_skipped = 0;
8234 : :
8235 [ + + ]: 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8236 : : {
8237 : 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8238 : 16384 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8239 : : bool buffer_already_dirty;
8240 : :
8241 [ - + ]: 16384 : CHECK_FOR_INTERRUPTS();
8242 : :
8243 : : /* An unlocked precheck should be safe and saves some cycles. */
8244 [ + + ]: 16384 : if ((buf_state & BM_VALID) == 0 ||
8245 [ + - ]: 27 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8246 : 16384 : continue;
8247 : :
8248 : : /* Make sure we can pin the buffer. */
8249 : 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
8250 : 0 : ReservePrivateRefCountEntry();
8251 : :
8252 : 0 : buf_state = LockBufHdr(desc);
8253 : :
8254 : : /* recheck, could have changed without the lock */
8255 [ # # ]: 0 : if ((buf_state & BM_VALID) == 0 ||
8256 [ # # ]: 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8257 : : {
8258 : 0 : UnlockBufHdr(desc);
8259 : 0 : continue;
8260 : : }
8261 : :
8262 [ # # ]: 0 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8263 : 0 : (*buffers_dirtied)++;
8264 [ # # ]: 0 : else if (buffer_already_dirty)
8265 : 0 : (*buffers_already_dirty)++;
8266 : : else
8267 : 0 : (*buffers_skipped)++;
8268 : : }
8269 : 1 : }
8270 : :
8271 : : /*
8272 : : * Try to mark all the shared buffers as dirty.
8273 : : *
8274 : : * This function is intended for testing/development use only! See
8275 : : * MarkDirtyUnpinnedBuffer().
8276 : : *
8277 : : * See MarkDirtyRelUnpinnedBuffers() above for details about the buffers_*
8278 : : * parameters.
8279 : : */
8280 : : void
8281 : 1 : MarkDirtyAllUnpinnedBuffers(int32 *buffers_dirtied,
8282 : : int32 *buffers_already_dirty,
8283 : : int32 *buffers_skipped)
8284 : : {
8285 : 1 : *buffers_dirtied = 0;
8286 : 1 : *buffers_already_dirty = 0;
8287 : 1 : *buffers_skipped = 0;
8288 : :
8289 [ + + ]: 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8290 : : {
8291 : 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8292 : : uint64 buf_state;
8293 : : bool buffer_already_dirty;
8294 : :
8295 [ - + ]: 16384 : CHECK_FOR_INTERRUPTS();
8296 : :
8297 : 16384 : buf_state = pg_atomic_read_u64(&desc->state);
8298 [ + + ]: 16384 : if (!(buf_state & BM_VALID))
8299 : 16349 : continue;
8300 : :
8301 : 35 : ResourceOwnerEnlarge(CurrentResourceOwner);
8302 : 35 : ReservePrivateRefCountEntry();
8303 : :
8304 : 35 : LockBufHdr(desc);
8305 : :
8306 [ + + ]: 35 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8307 : 16 : (*buffers_dirtied)++;
8308 [ + - ]: 19 : else if (buffer_already_dirty)
8309 : 19 : (*buffers_already_dirty)++;
8310 : : else
8311 : 0 : (*buffers_skipped)++;
8312 : : }
8313 : 1 : }
8314 : :
8315 : : /*
8316 : : * Generic implementation of the AIO handle staging callback for readv/writev
8317 : : * on local/shared buffers.
8318 : : *
8319 : : * Each readv/writev can target multiple buffers. The buffers have already
8320 : : * been registered with the IO handle.
8321 : : *
8322 : : * To make the IO ready for execution ("staging"), we need to ensure that the
8323 : : * targeted buffers are in an appropriate state while the IO is ongoing. For
8324 : : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
8325 : : * in this backend could lead to this backend's buffer pin being released as
8326 : : * part of error handling, which in turn could lead to the buffer being
8327 : : * replaced while IO is ongoing.
8328 : : */
8329 : : static pg_always_inline void
8330 : 1397631 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
8331 : : {
8332 : : uint64 *io_data;
8333 : : uint8 handle_data_len;
8334 : : PgAioWaitRef io_ref;
8335 : 1397631 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
8336 : :
8337 : 1397631 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8338 : :
8339 : 1397631 : pgaio_io_get_wref(ioh, &io_ref);
8340 : :
8341 : : /* iterate over all buffers affected by the vectored readv/writev */
8342 [ + + ]: 2980086 : for (int i = 0; i < handle_data_len; i++)
8343 : : {
8344 : 1582455 : Buffer buffer = (Buffer) io_data[i];
8345 : 1582455 : BufferDesc *buf_hdr = is_temp ?
8346 : 11063 : GetLocalBufferDescriptor(-buffer - 1)
8347 [ + + ]: 1582455 : : GetBufferDescriptor(buffer - 1);
8348 : : uint64 buf_state;
8349 : :
8350 : : /*
8351 : : * Check that all the buffers are actually ones that could conceivably
8352 : : * be done in one IO, i.e. are sequential. This is the last
8353 : : * buffer-aware code before IO is actually executed and confusion
8354 : : * about which buffers are targeted by IO can be hard to debug, making
8355 : : * it worth doing extra-paranoid checks.
8356 : : */
8357 [ + + ]: 1582455 : if (i == 0)
8358 : 1397631 : first = buf_hdr->tag;
8359 : : else
8360 : : {
8361 : : Assert(buf_hdr->tag.relNumber == first.relNumber);
8362 : : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
8363 : : }
8364 : :
8365 [ + + ]: 1582455 : if (is_temp)
8366 : 11063 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
8367 : : else
8368 : 1571392 : buf_state = LockBufHdr(buf_hdr);
8369 : :
8370 : : /* verify the buffer is in the expected state */
8371 : : Assert(buf_state & BM_TAG_VALID);
8372 : : if (is_write)
8373 : : {
8374 : : Assert(buf_state & BM_VALID);
8375 : : Assert(buf_state & BM_DIRTY);
8376 : : }
8377 : : else
8378 : : {
8379 : : Assert(!(buf_state & BM_VALID));
8380 : : Assert(!(buf_state & BM_DIRTY));
8381 : : }
8382 : :
8383 : : /* temp buffers don't use BM_IO_IN_PROGRESS */
8384 : 1582455 : if (!is_temp)
8385 : : Assert(buf_state & BM_IO_IN_PROGRESS);
8386 : :
8387 : : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
8388 : :
8389 : : /*
8390 : : * Reflect that the buffer is now owned by the AIO subsystem.
8391 : : *
8392 : : * For local buffers: This can't be done just via LocalRefCount, as
8393 : : * one might initially think, as this backend could error out while
8394 : : * AIO is still in progress, releasing all the pins by the backend
8395 : : * itself.
8396 : : *
8397 : : * This pin is released again in TerminateBufferIO().
8398 : : */
8399 : 1582455 : buf_hdr->io_wref = io_ref;
8400 : :
8401 [ + + ]: 1582455 : if (is_temp)
8402 : : {
8403 : 11063 : buf_state += BUF_REFCOUNT_ONE;
8404 : 11063 : pg_atomic_unlocked_write_u64(&buf_hdr->state, buf_state);
8405 : : }
8406 : : else
8407 : 1571392 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
8408 : :
8409 : : /*
8410 : : * Ensure the content lock that prevents buffer modifications while
8411 : : * the buffer is being written out is not released early due to an
8412 : : * error.
8413 : : */
8414 [ - + - - ]: 1582455 : if (is_write && !is_temp)
8415 : : {
8416 : : Assert(BufferLockHeldByMe(buf_hdr));
8417 : :
8418 : : /*
8419 : : * Lock is now owned by AIO subsystem.
8420 : : */
8421 : 0 : BufferLockDisown(buffer, buf_hdr);
8422 : : }
8423 : :
8424 : : /*
8425 : : * Stop tracking this buffer via the resowner - the AIO system now
8426 : : * keeps track.
8427 : : */
8428 [ + + ]: 1582455 : if (!is_temp)
8429 : 1571392 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
8430 : : }
8431 : 1397631 : }
8432 : :
8433 : : /*
8434 : : * Decode readv errors as encoded by buffer_readv_encode_error().
8435 : : */
8436 : : static inline void
8437 : 454 : buffer_readv_decode_error(PgAioResult result,
8438 : : bool *zeroed_any,
8439 : : bool *ignored_any,
8440 : : uint8 *zeroed_or_error_count,
8441 : : uint8 *checkfail_count,
8442 : : uint8 *first_off)
8443 : : {
8444 : 454 : uint32 rem_error = result.error_data;
8445 : :
8446 : : /* see static asserts in buffer_readv_encode_error */
8447 : : #define READV_COUNT_BITS 7
8448 : : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
8449 : :
8450 : 454 : *zeroed_any = rem_error & 1;
8451 : 454 : rem_error >>= 1;
8452 : :
8453 : 454 : *ignored_any = rem_error & 1;
8454 : 454 : rem_error >>= 1;
8455 : :
8456 : 454 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
8457 : 454 : rem_error >>= READV_COUNT_BITS;
8458 : :
8459 : 454 : *checkfail_count = rem_error & READV_COUNT_MASK;
8460 : 454 : rem_error >>= READV_COUNT_BITS;
8461 : :
8462 : 454 : *first_off = rem_error & READV_COUNT_MASK;
8463 : 454 : rem_error >>= READV_COUNT_BITS;
8464 : 454 : }
8465 : :
8466 : : /*
8467 : : * Helper to encode errors for buffer_readv_complete()
8468 : : *
8469 : : * Errors are encoded as follows:
8470 : : * - bit 0 indicates whether any page was zeroed (1) or not (0)
8471 : : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
8472 : : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
8473 : : * - next READV_COUNT_BITS bits indicate the number of checksum failures
8474 : : * - next READV_COUNT_BITS bits indicate the first offset of the first page
8475 : : * that was errored or zeroed or, if no errors/zeroes, the first ignored
8476 : : * checksum
8477 : : */
8478 : : static inline void
8479 : 194 : buffer_readv_encode_error(PgAioResult *result,
8480 : : bool is_temp,
8481 : : bool zeroed_any,
8482 : : bool ignored_any,
8483 : : uint8 error_count,
8484 : : uint8 zeroed_count,
8485 : : uint8 checkfail_count,
8486 : : uint8 first_error_off,
8487 : : uint8 first_zeroed_off,
8488 : : uint8 first_ignored_off)
8489 : : {
8490 : :
8491 : 194 : uint8 shift = 0;
8492 [ + + ]: 194 : uint8 zeroed_or_error_count =
8493 : : error_count > 0 ? error_count : zeroed_count;
8494 : : uint8 first_off;
8495 : :
8496 : : StaticAssertDecl(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
8497 : : "PG_IOV_MAX is bigger than reserved space for error data");
8498 : : StaticAssertDecl((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
8499 : : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
8500 : :
8501 : : /*
8502 : : * We only have space to encode one offset - but luckily that's good
8503 : : * enough. If there is an error, the error is the interesting offset, same
8504 : : * with a zeroed buffer vs an ignored buffer.
8505 : : */
8506 [ + + ]: 194 : if (error_count > 0)
8507 : 94 : first_off = first_error_off;
8508 [ + + ]: 100 : else if (zeroed_count > 0)
8509 : 82 : first_off = first_zeroed_off;
8510 : : else
8511 : 18 : first_off = first_ignored_off;
8512 : :
8513 : : Assert(!zeroed_any || error_count == 0);
8514 : :
8515 : 194 : result->error_data = 0;
8516 : :
8517 : 194 : result->error_data |= zeroed_any << shift;
8518 : 194 : shift += 1;
8519 : :
8520 : 194 : result->error_data |= ignored_any << shift;
8521 : 194 : shift += 1;
8522 : :
8523 : 194 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
8524 : 194 : shift += READV_COUNT_BITS;
8525 : :
8526 : 194 : result->error_data |= ((uint32) checkfail_count) << shift;
8527 : 194 : shift += READV_COUNT_BITS;
8528 : :
8529 : 194 : result->error_data |= ((uint32) first_off) << shift;
8530 : 194 : shift += READV_COUNT_BITS;
8531 : :
8532 [ + + ]: 194 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
8533 : : PGAIO_HCB_SHARED_BUFFER_READV;
8534 : :
8535 [ + + ]: 194 : if (error_count > 0)
8536 : 94 : result->status = PGAIO_RS_ERROR;
8537 : : else
8538 : 100 : result->status = PGAIO_RS_WARNING;
8539 : :
8540 : : /*
8541 : : * The encoding is complicated enough to warrant cross-checking it against
8542 : : * the decode function.
8543 : : */
8544 : : #ifdef USE_ASSERT_CHECKING
8545 : : {
8546 : : bool zeroed_any_2,
8547 : : ignored_any_2;
8548 : : uint8 zeroed_or_error_count_2,
8549 : : checkfail_count_2,
8550 : : first_off_2;
8551 : :
8552 : : buffer_readv_decode_error(*result,
8553 : : &zeroed_any_2, &ignored_any_2,
8554 : : &zeroed_or_error_count_2,
8555 : : &checkfail_count_2,
8556 : : &first_off_2);
8557 : : Assert(zeroed_any == zeroed_any_2);
8558 : : Assert(ignored_any == ignored_any_2);
8559 : : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
8560 : : Assert(checkfail_count == checkfail_count_2);
8561 : : Assert(first_off == first_off_2);
8562 : : }
8563 : : #endif
8564 : :
8565 : : #undef READV_COUNT_BITS
8566 : : #undef READV_COUNT_MASK
8567 : 194 : }
8568 : :
8569 : : /*
8570 : : * Helper for AIO readv completion callbacks, supporting both shared and temp
8571 : : * buffers. Gets called once for each buffer in a multi-page read.
8572 : : */
8573 : : static pg_always_inline void
8574 : 1408270 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
8575 : : uint8 flags, bool failed, bool is_temp,
8576 : : bool *buffer_invalid,
8577 : : bool *failed_checksum,
8578 : : bool *ignored_checksum,
8579 : : bool *zeroed_buffer)
8580 : : {
8581 : 1408270 : BufferDesc *buf_hdr = is_temp ?
8582 : 11063 : GetLocalBufferDescriptor(-buffer - 1)
8583 [ + + ]: 1408270 : : GetBufferDescriptor(buffer - 1);
8584 : 1408270 : BufferTag tag = buf_hdr->tag;
8585 : 1408270 : char *bufdata = BufferGetBlock(buffer);
8586 : : uint64 set_flag_bits;
8587 : : int piv_flags;
8588 : :
8589 : : /* check that the buffer is in the expected state for a read */
8590 : : #ifdef USE_ASSERT_CHECKING
8591 : : {
8592 : : uint64 buf_state = pg_atomic_read_u64(&buf_hdr->state);
8593 : :
8594 : : Assert(buf_state & BM_TAG_VALID);
8595 : : Assert(!(buf_state & BM_VALID));
8596 : : /* temp buffers don't use BM_IO_IN_PROGRESS */
8597 : : if (!is_temp)
8598 : : Assert(buf_state & BM_IO_IN_PROGRESS);
8599 : : Assert(!(buf_state & BM_DIRTY));
8600 : : }
8601 : : #endif
8602 : :
8603 : 1408270 : *buffer_invalid = false;
8604 : 1408270 : *failed_checksum = false;
8605 : 1408270 : *ignored_checksum = false;
8606 : 1408270 : *zeroed_buffer = false;
8607 : :
8608 : : /*
8609 : : * We ask PageIsVerified() to only log the message about checksum errors,
8610 : : * as the completion might be run in any backend (or IO workers). We will
8611 : : * report checksum errors in buffer_readv_report().
8612 : : */
8613 : 1408270 : piv_flags = PIV_LOG_LOG;
8614 : :
8615 : : /* the local zero_damaged_pages may differ from the definer's */
8616 [ + + ]: 1408270 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
8617 : 38 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
8618 : :
8619 : : /*
8620 : : * If the buffers are marked for zero on error, we want to log that in
8621 : : * case of a checksum failure.
8622 : : */
8623 [ + + ]: 1408270 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8624 : 46200 : piv_flags |= PIV_ZERO_BUFFERS_ON_ERROR;
8625 : :
8626 : : /* Check for garbage data. */
8627 [ + + ]: 1408270 : if (!failed)
8628 : : {
8629 : : /*
8630 : : * If the buffer is not currently pinned by this backend, e.g. because
8631 : : * we're completing this IO after an error, the buffer data will have
8632 : : * been marked as inaccessible when the buffer was unpinned. The AIO
8633 : : * subsystem holds a pin, but that doesn't prevent the buffer from
8634 : : * having been marked as inaccessible. The completion might also be
8635 : : * executed in a different process.
8636 : : */
8637 : : #ifdef USE_VALGRIND
8638 : : if (!BufferIsPinned(buffer))
8639 : : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
8640 : : #endif
8641 : :
8642 [ + + ]: 1407498 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
8643 : : failed_checksum))
8644 : : {
8645 [ + + ]: 97 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8646 : : {
8647 : 47 : memset(bufdata, 0, BLCKSZ);
8648 : 47 : *zeroed_buffer = true;
8649 : : }
8650 : : else
8651 : : {
8652 : 50 : *buffer_invalid = true;
8653 : : /* mark buffer as having failed */
8654 : 50 : failed = true;
8655 : : }
8656 : : }
8657 [ + + ]: 1407401 : else if (*failed_checksum)
8658 : 12 : *ignored_checksum = true;
8659 : :
8660 : : /* undo what we did above */
8661 : : #ifdef USE_VALGRIND
8662 : : if (!BufferIsPinned(buffer))
8663 : : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
8664 : : #endif
8665 : :
8666 : : /*
8667 : : * Immediately log a message about the invalid page, but only to the
8668 : : * server log. The reason to do so immediately is that this may be
8669 : : * executed in a different backend than the one that originated the
8670 : : * request. The reason to do so immediately is that the originator
8671 : : * might not process the query result immediately (because it is busy
8672 : : * doing another part of query processing) or at all (e.g. if it was
8673 : : * cancelled or errored out due to another IO also failing). The
8674 : : * definer of the IO will emit an ERROR or WARNING when processing the
8675 : : * IO's results
8676 : : *
8677 : : * To avoid duplicating the code to emit these log messages, we reuse
8678 : : * buffer_readv_report().
8679 : : */
8680 [ + + + + : 1407498 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
+ + ]
8681 : : {
8682 : 109 : PgAioResult result_one = {0};
8683 : :
8684 : 109 : buffer_readv_encode_error(&result_one, is_temp,
8685 : 109 : *zeroed_buffer,
8686 : 109 : *ignored_checksum,
8687 : 109 : *buffer_invalid,
8688 : 109 : *zeroed_buffer ? 1 : 0,
8689 : 109 : *failed_checksum ? 1 : 0,
8690 : : buf_off, buf_off, buf_off);
8691 : 109 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
8692 : : }
8693 : : }
8694 : :
8695 : : /* Terminate I/O and set BM_VALID. */
8696 [ + + ]: 1408270 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
8697 [ + + ]: 1408270 : if (is_temp)
8698 : 11063 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
8699 : : else
8700 : 1397207 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
8701 : :
8702 : : /*
8703 : : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
8704 : : * callback may not be executed in the same backend that called
8705 : : * BUFFER_READ_START. The alternative would be to defer calling the
8706 : : * tracepoint to a later point (e.g. the local completion callback for
8707 : : * shared buffer reads), which seems even less helpful.
8708 : : */
8709 : : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
8710 : : tag.blockNum,
8711 : : tag.spcOid,
8712 : : tag.dbOid,
8713 : : tag.relNumber,
8714 : : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
8715 : : false);
8716 : 1408270 : }
8717 : :
8718 : : /*
8719 : : * Perform completion handling of a single AIO read. This read may cover
8720 : : * multiple blocks / buffers.
8721 : : *
8722 : : * Shared between shared and local buffers, to reduce code duplication.
8723 : : */
8724 : : static pg_always_inline PgAioResult
8725 : 1259745 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8726 : : uint8 cb_data, bool is_temp)
8727 : : {
8728 : 1259745 : PgAioResult result = prior_result;
8729 : 1259745 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8730 : 1259745 : uint8 first_error_off = 0;
8731 : 1259745 : uint8 first_zeroed_off = 0;
8732 : 1259745 : uint8 first_ignored_off = 0;
8733 : 1259745 : uint8 error_count = 0;
8734 : 1259745 : uint8 zeroed_count = 0;
8735 : 1259745 : uint8 ignored_count = 0;
8736 : 1259745 : uint8 checkfail_count = 0;
8737 : : uint64 *io_data;
8738 : : uint8 handle_data_len;
8739 : :
8740 : : if (is_temp)
8741 : : {
8742 : : Assert(td->smgr.is_temp);
8743 : : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
8744 : : }
8745 : : else
8746 : : Assert(!td->smgr.is_temp);
8747 : :
8748 : : /*
8749 : : * Iterate over all the buffers affected by this IO and call the
8750 : : * per-buffer completion function for each buffer.
8751 : : */
8752 : 1259745 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8753 [ + + ]: 2668015 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
8754 : : {
8755 : 1408270 : Buffer buf = io_data[buf_off];
8756 : : bool failed;
8757 : 1408270 : bool failed_verification = false;
8758 : 1408270 : bool failed_checksum = false;
8759 : 1408270 : bool zeroed_buffer = false;
8760 : 1408270 : bool ignored_checksum = false;
8761 : :
8762 : : Assert(BufferIsValid(buf));
8763 : :
8764 : : /*
8765 : : * If the entire I/O failed on a lower-level, each buffer needs to be
8766 : : * marked as failed. In case of a partial read, the first few buffers
8767 : : * may be ok.
8768 : : */
8769 : 1408270 : failed =
8770 : 1408270 : prior_result.status == PGAIO_RS_ERROR
8771 [ + + + + ]: 1408270 : || prior_result.result <= buf_off;
8772 : :
8773 : 1408270 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
8774 : : &failed_verification,
8775 : : &failed_checksum,
8776 : : &ignored_checksum,
8777 : : &zeroed_buffer);
8778 : :
8779 : : /*
8780 : : * Track information about the number of different kinds of error
8781 : : * conditions across all pages, as there can be multiple pages failing
8782 : : * verification as part of one IO.
8783 : : */
8784 [ + + + - : 1408270 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
+ + ]
8785 : 44 : first_error_off = buf_off;
8786 [ + + + + ]: 1408270 : if (zeroed_buffer && zeroed_count++ == 0)
8787 : 35 : first_zeroed_off = buf_off;
8788 [ + + + + ]: 1408270 : if (ignored_checksum && ignored_count++ == 0)
8789 : 10 : first_ignored_off = buf_off;
8790 [ + + ]: 1408270 : if (failed_checksum)
8791 : 33 : checkfail_count++;
8792 : : }
8793 : :
8794 : : /*
8795 : : * If the smgr read succeeded [partially] and page verification failed for
8796 : : * some of the pages, adjust the IO's result state appropriately.
8797 : : */
8798 [ + + + + ]: 1259745 : if (prior_result.status != PGAIO_RS_ERROR &&
8799 [ + + + + ]: 1259690 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
8800 : : {
8801 : 85 : buffer_readv_encode_error(&result, is_temp,
8802 : : zeroed_count > 0, ignored_count > 0,
8803 : : error_count, zeroed_count, checkfail_count,
8804 : : first_error_off, first_zeroed_off,
8805 : : first_ignored_off);
8806 : 85 : pgaio_result_report(result, td, DEBUG1);
8807 : : }
8808 : :
8809 : : /*
8810 : : * For shared relations this reporting is done in
8811 : : * shared_buffer_readv_complete_local().
8812 : : */
8813 [ + + + + ]: 1259745 : if (is_temp && checkfail_count > 0)
8814 : 2 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8815 : : checkfail_count);
8816 : :
8817 : 1259745 : return result;
8818 : : }
8819 : :
8820 : : /*
8821 : : * AIO error reporting callback for aio_shared_buffer_readv_cb and
8822 : : * aio_local_buffer_readv_cb.
8823 : : *
8824 : : * The error is encoded / decoded in buffer_readv_encode_error() /
8825 : : * buffer_readv_decode_error().
8826 : : */
8827 : : static void
8828 : 275 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
8829 : : int elevel)
8830 : : {
8831 : 275 : int nblocks = td->smgr.nblocks;
8832 : 275 : BlockNumber first = td->smgr.blockNum;
8833 : 275 : BlockNumber last = first + nblocks - 1;
8834 : 275 : ProcNumber errProc =
8835 [ + + ]: 275 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
8836 : : RelPathStr rpath =
8837 : 275 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
8838 : : bool zeroed_any,
8839 : : ignored_any;
8840 : : uint8 zeroed_or_error_count,
8841 : : checkfail_count,
8842 : : first_off;
8843 : : uint8 affected_count;
8844 : : const char *msg_one,
8845 : : *msg_mult,
8846 : : *det_mult,
8847 : : *hint_mult;
8848 : :
8849 : 275 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
8850 : : &zeroed_or_error_count,
8851 : : &checkfail_count,
8852 : : &first_off);
8853 : :
8854 : : /*
8855 : : * Treat a read that had both zeroed buffers *and* ignored checksums as a
8856 : : * special case, it's too irregular to be emitted the same way as the
8857 : : * other cases.
8858 : : */
8859 [ + + + + ]: 275 : if (zeroed_any && ignored_any)
8860 : : {
8861 : : Assert(zeroed_any && ignored_any);
8862 : : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
8863 : : Assert(result.status != PGAIO_RS_ERROR);
8864 : 4 : affected_count = zeroed_or_error_count;
8865 : :
8866 [ + - + - ]: 4 : ereport(elevel,
8867 : : errcode(ERRCODE_DATA_CORRUPTED),
8868 : : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
8869 : : affected_count, checkfail_count, first, last, rpath.str),
8870 : : affected_count > 1 ?
8871 : : errdetail("Block %u held the first zeroed page.",
8872 : : first + first_off) : 0,
8873 : : errhint_plural("See server log for details about the other %d invalid block.",
8874 : : "See server log for details about the other %d invalid blocks.",
8875 : : affected_count + checkfail_count - 1,
8876 : : affected_count + checkfail_count - 1));
8877 : 4 : return;
8878 : : }
8879 : :
8880 : : /*
8881 : : * The other messages are highly repetitive. To avoid duplicating a long
8882 : : * and complicated ereport(), gather the translated format strings
8883 : : * separately and then do one common ereport.
8884 : : */
8885 [ + + ]: 271 : if (result.status == PGAIO_RS_ERROR)
8886 : : {
8887 : : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
8888 : 136 : affected_count = zeroed_or_error_count;
8889 : 136 : msg_one = _("invalid page in block %u of relation \"%s\"");
8890 : 136 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
8891 : 136 : det_mult = _("Block %u held the first invalid page.");
8892 : 136 : hint_mult = _("See server log for the other %u invalid block(s).");
8893 : : }
8894 [ + + + - ]: 135 : else if (zeroed_any && !ignored_any)
8895 : : {
8896 : 111 : affected_count = zeroed_or_error_count;
8897 : 111 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
8898 : 111 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
8899 : 111 : det_mult = _("Block %u held the first zeroed page.");
8900 : 111 : hint_mult = _("See server log for the other %u zeroed block(s).");
8901 : : }
8902 [ + - + - ]: 24 : else if (!zeroed_any && ignored_any)
8903 : : {
8904 : 24 : affected_count = checkfail_count;
8905 : 24 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
8906 : 24 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
8907 : 24 : det_mult = _("Block %u held the first ignored page.");
8908 : 24 : hint_mult = _("See server log for the other %u ignored block(s).");
8909 : : }
8910 : : else
8911 : 0 : pg_unreachable();
8912 : :
8913 [ + + + + : 271 : ereport(elevel,
+ + + + ]
8914 : : errcode(ERRCODE_DATA_CORRUPTED),
8915 : : affected_count == 1 ?
8916 : : errmsg_internal(msg_one, first + first_off, rpath.str) :
8917 : : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
8918 : : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
8919 : : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
8920 : : }
8921 : :
8922 : : static void
8923 : 1394784 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8924 : : {
8925 : 1394784 : buffer_stage_common(ioh, false, false);
8926 : 1394784 : }
8927 : :
8928 : : static PgAioResult
8929 : 1256898 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8930 : : uint8 cb_data)
8931 : : {
8932 : 1256898 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
8933 : : }
8934 : :
8935 : : /*
8936 : : * We need a backend-local completion callback for shared buffers, to be able
8937 : : * to report checksum errors correctly. Unfortunately that can only safely
8938 : : * happen if the reporting backend has previously called
8939 : : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
8940 : : * the backend that started the IO. Hence this callback.
8941 : : */
8942 : : static PgAioResult
8943 : 1394784 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
8944 : : uint8 cb_data)
8945 : : {
8946 : : bool zeroed_any,
8947 : : ignored_any;
8948 : : uint8 zeroed_or_error_count,
8949 : : checkfail_count,
8950 : : first_off;
8951 : :
8952 [ + + ]: 1394784 : if (prior_result.status == PGAIO_RS_OK)
8953 : 1394605 : return prior_result;
8954 : :
8955 : 179 : buffer_readv_decode_error(prior_result,
8956 : : &zeroed_any,
8957 : : &ignored_any,
8958 : : &zeroed_or_error_count,
8959 : : &checkfail_count,
8960 : : &first_off);
8961 : :
8962 [ + + ]: 179 : if (checkfail_count)
8963 : : {
8964 : 25 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8965 : :
8966 : 25 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8967 : : checkfail_count);
8968 : : }
8969 : :
8970 : 179 : return prior_result;
8971 : : }
8972 : :
8973 : : static void
8974 : 2847 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8975 : : {
8976 : 2847 : buffer_stage_common(ioh, false, true);
8977 : 2847 : }
8978 : :
8979 : : static PgAioResult
8980 : 2847 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8981 : : uint8 cb_data)
8982 : : {
8983 : 2847 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
8984 : : }
8985 : :
8986 : : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8987 : : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
8988 : : .stage = shared_buffer_readv_stage,
8989 : : .complete_shared = shared_buffer_readv_complete,
8990 : : /* need a local callback to report checksum failures */
8991 : : .complete_local = shared_buffer_readv_complete_local,
8992 : : .report = buffer_readv_report,
8993 : : };
8994 : :
8995 : : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8996 : : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
8997 : : .stage = local_buffer_readv_stage,
8998 : :
8999 : : /*
9000 : : * Note that this, in contrast to the shared_buffers case, uses
9001 : : * complete_local, as only the issuing backend has access to the required
9002 : : * datastructures. This is important in case the IO completion may be
9003 : : * consumed incidentally by another backend.
9004 : : */
9005 : : .complete_local = local_buffer_readv_complete,
9006 : : .report = buffer_readv_report,
9007 : : };
|
